mirror of
https://github.com/open-jarvis/OpenJarvis.git
synced 2026-08-14 08:52:06 +00:00
Compare commits
5
Commits
| Author | SHA1 | Date | |
|---|---|---|---|
|
|
08fe3a6c98 | ||
|
|
f338d64b21 | ||
|
|
6862b50da8 | ||
|
|
08916ec468 | ||
|
|
28ef7e5cb5 |
@@ -130,3 +130,10 @@ learning.db
|
||||
minion_logs/
|
||||
*.oj-debug.json
|
||||
oj-debug.*.json
|
||||
|
||||
# Generated orchestrator SFT/eval data artifacts
|
||||
data/
|
||||
|
||||
# Training run artifacts (model checkpoints + W&B run dirs — regenerated, large)
|
||||
checkpoints/
|
||||
wandb/
|
||||
|
||||
@@ -0,0 +1,74 @@
|
||||
# LiteLLM proxy config for OpenJarvis ToolOrchestra expert spend tracking.
|
||||
# Launch: litellm --config litellm/config.yaml --port 4000 --host 127.0.0.1
|
||||
# Keys are read from process env (source OpenJarvis/.env first).
|
||||
# No Postgres on this box (docker socket denied) -> admin UI/virtual keys are
|
||||
# disabled; spend is written to litellm/logs/spend.jsonl by spend_logger.py.
|
||||
|
||||
model_list:
|
||||
# ---- OpenAI ----
|
||||
- model_name: gpt-5.5
|
||||
litellm_params:
|
||||
model: openai/gpt-5.5
|
||||
api_key: os.environ/OPENAI_API_KEY
|
||||
- model_name: gpt-5-mini
|
||||
litellm_params:
|
||||
model: openai/gpt-5-mini
|
||||
api_key: os.environ/OPENAI_API_KEY
|
||||
- model_name: gpt-4o
|
||||
litellm_params:
|
||||
model: openai/gpt-4o
|
||||
api_key: os.environ/OPENAI_API_KEY
|
||||
|
||||
# ---- Anthropic ----
|
||||
- model_name: claude-opus-4-8
|
||||
litellm_params:
|
||||
model: anthropic/claude-opus-4-8
|
||||
api_key: os.environ/ANTHROPIC_API_KEY
|
||||
- model_name: claude-sonnet-4-6
|
||||
litellm_params:
|
||||
model: anthropic/claude-sonnet-4-6
|
||||
api_key: os.environ/ANTHROPIC_API_KEY
|
||||
|
||||
# ---- Gemini ----
|
||||
- model_name: gemini-2.5-pro
|
||||
litellm_params:
|
||||
model: gemini/gemini-2.5-pro
|
||||
api_key: os.environ/GEMINI_API_KEY
|
||||
- model_name: gemini-2.5-flash
|
||||
litellm_params:
|
||||
model: gemini/gemini-2.5-flash
|
||||
api_key: os.environ/GEMINI_API_KEY
|
||||
|
||||
# ---- OpenRouter (Qwen coder etc.) ----
|
||||
- model_name: qwen-coder
|
||||
litellm_params:
|
||||
model: openrouter/qwen/qwen-2.5-coder-32b-instruct
|
||||
api_key: os.environ/OPENROUTER_API_KEY
|
||||
- model_name: openrouter/*
|
||||
litellm_params:
|
||||
model: openrouter/*
|
||||
api_key: os.environ/OPENROUTER_API_KEY
|
||||
|
||||
# ---- Local vLLM (self-hosted students / experts) ----
|
||||
- model_name: qwen3.5-9b-local
|
||||
litellm_params:
|
||||
model: openai/Qwen/Qwen3.5-9B
|
||||
api_base: http://localhost:8011/v1
|
||||
api_key: EMPTY
|
||||
input_cost_per_token: 0.0
|
||||
output_cost_per_token: 0.0
|
||||
- model_name: qwen3.6-27b-local
|
||||
litellm_params:
|
||||
model: openai/Qwen/Qwen3.6-27B-FP8
|
||||
api_base: http://localhost:8002/v1
|
||||
api_key: EMPTY
|
||||
input_cost_per_token: 0.0
|
||||
output_cost_per_token: 0.0
|
||||
|
||||
litellm_settings:
|
||||
# JSONL spend logger (Postgres-free spend tracking).
|
||||
callbacks: spend_logger.proxy_handler_instance
|
||||
drop_params: true
|
||||
|
||||
general_settings:
|
||||
master_key: os.environ/LITELLM_MASTER_KEY
|
||||
@@ -0,0 +1,87 @@
|
||||
"""Custom LiteLLM callback that appends per-request cost/usage to a JSONL file.
|
||||
|
||||
This is the spend-tracking fallback used when no Postgres is available for the
|
||||
LiteLLM admin UI. Every successful (and failed) proxy request writes one line to
|
||||
$OJ_LITELLM_SPEND_LOG. Aggregate with `litellm/spend_report.py`.
|
||||
"""
|
||||
|
||||
from __future__ import annotations
|
||||
|
||||
import datetime
|
||||
import json
|
||||
import os
|
||||
import threading
|
||||
|
||||
from litellm.integrations.custom_logger import CustomLogger
|
||||
|
||||
LOG_PATH = os.environ.get(
|
||||
"OJ_LITELLM_SPEND_LOG",
|
||||
os.path.join(os.path.dirname(os.path.abspath(__file__)), "logs", "spend.jsonl"),
|
||||
)
|
||||
os.makedirs(os.path.dirname(LOG_PATH), exist_ok=True)
|
||||
|
||||
_lock = threading.Lock()
|
||||
|
||||
|
||||
def _write(kwargs, response_obj, start_time, end_time, status: str) -> None:
|
||||
try:
|
||||
cost = kwargs.get("response_cost")
|
||||
litellm_params = kwargs.get("litellm_params") or {}
|
||||
metadata = litellm_params.get("metadata") or {}
|
||||
usage = None
|
||||
try:
|
||||
usage = getattr(response_obj, "usage", None) or (
|
||||
response_obj.get("usage") if isinstance(response_obj, dict) else None
|
||||
)
|
||||
except Exception:
|
||||
usage = None
|
||||
|
||||
def _u(field):
|
||||
if usage is None:
|
||||
return None
|
||||
return getattr(usage, field, None) if not isinstance(usage, dict) else usage.get(field)
|
||||
|
||||
dur = None
|
||||
try:
|
||||
if start_time and end_time:
|
||||
dur = (end_time - start_time).total_seconds()
|
||||
except Exception:
|
||||
dur = None
|
||||
|
||||
rec = {
|
||||
"ts": datetime.datetime.utcnow().isoformat() + "Z",
|
||||
"status": status,
|
||||
"model": kwargs.get("model"),
|
||||
"model_group": metadata.get("model_group"),
|
||||
"call_type": kwargs.get("call_type"),
|
||||
"cost_usd": cost,
|
||||
"prompt_tokens": _u("prompt_tokens"),
|
||||
"completion_tokens": _u("completion_tokens"),
|
||||
"total_tokens": _u("total_tokens"),
|
||||
"duration_s": dur,
|
||||
"key_alias": metadata.get("user_api_key_alias"),
|
||||
"exception": str(kwargs.get("exception")) if status == "failure" else None,
|
||||
}
|
||||
line = json.dumps(rec)
|
||||
with _lock:
|
||||
with open(LOG_PATH, "a") as f:
|
||||
f.write(line + "\n")
|
||||
except Exception as exc: # never let logging break a request
|
||||
print(f"[spend_logger] error: {exc}")
|
||||
|
||||
|
||||
class SpendLogger(CustomLogger):
|
||||
def log_success_event(self, kwargs, response_obj, start_time, end_time):
|
||||
_write(kwargs, response_obj, start_time, end_time, "success")
|
||||
|
||||
async def async_log_success_event(self, kwargs, response_obj, start_time, end_time):
|
||||
_write(kwargs, response_obj, start_time, end_time, "success")
|
||||
|
||||
def log_failure_event(self, kwargs, response_obj, start_time, end_time):
|
||||
_write(kwargs, response_obj, start_time, end_time, "failure")
|
||||
|
||||
async def async_log_failure_event(self, kwargs, response_obj, start_time, end_time):
|
||||
_write(kwargs, response_obj, start_time, end_time, "failure")
|
||||
|
||||
|
||||
proxy_handler_instance = SpendLogger()
|
||||
@@ -0,0 +1,43 @@
|
||||
#!/usr/bin/env python3
|
||||
"""Aggregate litellm/logs/spend.jsonl into a per-model spend summary.
|
||||
|
||||
Postgres-free stand-in for the LiteLLM admin UI. Usage:
|
||||
.venv/bin/python litellm/spend_report.py [path-to-spend.jsonl]
|
||||
"""
|
||||
import json
|
||||
import os
|
||||
import sys
|
||||
from collections import defaultdict
|
||||
|
||||
path = sys.argv[1] if len(sys.argv) > 1 else os.environ.get(
|
||||
"OJ_LITELLM_SPEND_LOG",
|
||||
os.path.join(os.path.dirname(os.path.abspath(__file__)), "logs", "spend.jsonl"),
|
||||
)
|
||||
|
||||
agg = defaultdict(lambda: {"calls": 0, "ok": 0, "fail": 0, "cost": 0.0,
|
||||
"ptok": 0, "ctok": 0})
|
||||
total_cost = 0.0
|
||||
with open(path) as f:
|
||||
for line in f:
|
||||
line = line.strip()
|
||||
if not line:
|
||||
continue
|
||||
r = json.loads(line)
|
||||
g = r.get("model_group") or r.get("model") or "?"
|
||||
a = agg[g]
|
||||
a["calls"] += 1
|
||||
a["ok"] += 1 if r.get("status") == "success" else 0
|
||||
a["fail"] += 1 if r.get("status") == "failure" else 0
|
||||
a["cost"] += float(r.get("cost_usd") or 0)
|
||||
a["ptok"] += int(r.get("prompt_tokens") or 0)
|
||||
a["ctok"] += int(r.get("completion_tokens") or 0)
|
||||
total_cost += float(r.get("cost_usd") or 0)
|
||||
|
||||
hdr = f"{'model':<24}{'calls':>7}{'ok':>5}{'fail':>6}{'p_tok':>9}{'c_tok':>9}{'cost_usd':>12}"
|
||||
print(hdr)
|
||||
print("-" * len(hdr))
|
||||
for g, a in sorted(agg.items(), key=lambda kv: -kv[1]["cost"]):
|
||||
print(f"{g:<24}{a['calls']:>7}{a['ok']:>5}{a['fail']:>6}"
|
||||
f"{a['ptok']:>9}{a['ctok']:>9}{a['cost']:>12.6f}")
|
||||
print("-" * len(hdr))
|
||||
print(f"{'TOTAL':<24}{'':>7}{'':>5}{'':>6}{'':>9}{'':>9}{total_cost:>12.6f}")
|
||||
@@ -28,6 +28,7 @@ classifiers = [
|
||||
"Topic :: Scientific/Engineering :: Artificial Intelligence",
|
||||
]
|
||||
dependencies = [
|
||||
"braintrust>=0.0.150",
|
||||
"click>=8",
|
||||
"datasets>=4.5.0",
|
||||
"ddgs>=9.11.4",
|
||||
|
||||
@@ -0,0 +1,319 @@
|
||||
#!/usr/bin/env python
|
||||
"""Build orchestrator SFT cold-start data by **base Qwen3-8B self-sampling** (v1).
|
||||
|
||||
The orchestrator *is* the local Qwen3-8B. We roll the base (un-trained) model out
|
||||
over the reasoning-SFT task pool (``load_sft_tasks``: GeneralThought + OpenThoughts
|
||||
code/math/science), verify each trajectory's final answer against the gold
|
||||
(``verify.make_verifier`` — math/code checkers + Gemini judge fallback), and keep
|
||||
the cheapest-correct trajectory per task. Those passing trajectories become the
|
||||
``conversations`` JSONL the SFT trainer consumes — i.e. the model learns from its
|
||||
own successful rollouts (rejection-sampling cold-start, STaR-style).
|
||||
|
||||
This is **v1** (base self-sampling): point ``--orchestrator-endpoint`` at the vLLM
|
||||
server hosting the *base* ``Qwen/Qwen3-8B``.
|
||||
|
||||
For **v2**, train on the v1 data, serve the v1 checkpoint with vLLM, then re-run
|
||||
this exact driver with ``--orchestrator-endpoint`` pointed at the v1 checkpoint's
|
||||
endpoint, ``--orchestrator-model`` set to the checkpoint name, and
|
||||
``--samples-per-task 8`` — the (now stronger) policy self-samples a v2 dataset.
|
||||
|
||||
The math/coder specialist tools are only wired when ``--math-endpoint`` /
|
||||
``--coder-endpoint`` are passed (those are separately-served vLLM specialists);
|
||||
otherwise they're omitted so the orchestrator only sees the tools it can actually
|
||||
call.
|
||||
|
||||
Example (v1, base self-sampling):
|
||||
.venv/bin/python scripts/orchestrator/build_orchestrator_sft.py \
|
||||
--out data/orchestrator_sft_v1.jsonl \
|
||||
--orchestrator-endpoint http://localhost:8001/v1 \
|
||||
--orchestrator-model qwen3-8b \
|
||||
--samples-per-task 8 --max-keep-per-task 1
|
||||
|
||||
Example (v2, re-point at the v1 checkpoint):
|
||||
.venv/bin/python scripts/orchestrator/build_orchestrator_sft.py \
|
||||
--out data/orchestrator_sft_v2.jsonl \
|
||||
--orchestrator-endpoint http://localhost:8010/v1 \
|
||||
--orchestrator-model orchestrator-sft-v1 \
|
||||
--samples-per-task 8
|
||||
"""
|
||||
|
||||
from __future__ import annotations
|
||||
|
||||
import argparse
|
||||
import json
|
||||
import logging
|
||||
import os
|
||||
import sys
|
||||
import time
|
||||
from pathlib import Path
|
||||
from typing import Optional
|
||||
|
||||
from openjarvis.agents.hybrid.expert_registry import orchestrator_catalog
|
||||
from openjarvis.agents.hybrid.toolorchestra.rollout import run_unified_rollout
|
||||
from openjarvis.agents.hybrid.toolorchestra.unified import (
|
||||
make_call_orchestrator,
|
||||
make_dispatch,
|
||||
)
|
||||
from openjarvis.learning.intelligence.orchestrator.sft_data.datasets import (
|
||||
load_sft_tasks,
|
||||
)
|
||||
from openjarvis.learning.intelligence.orchestrator.sft_data.reject_sample import (
|
||||
generate_sft_dataset,
|
||||
)
|
||||
from openjarvis.learning.intelligence.orchestrator.sft_data.verify import make_verifier
|
||||
|
||||
|
||||
def main(argv: Optional[list[str]] = None) -> int:
|
||||
p = argparse.ArgumentParser(
|
||||
description=__doc__, formatter_class=argparse.RawDescriptionHelpFormatter
|
||||
)
|
||||
# Default output is timestamped (e.g. data/orchestrator_sft_06-22-1230am.jsonl)
|
||||
# so every run lands in its own file and concurrent runs never collide. Pass
|
||||
# --out explicitly to override (sharded runs do this, one file per shard).
|
||||
p.add_argument("--out", default=None,
|
||||
help="Output JSONL. Default: data/orchestrator_sft_<MM-DD-HHMMam>.jsonl")
|
||||
# The orchestrator == the local Qwen3-8B self-sampling over its own rollouts.
|
||||
p.add_argument("--orchestrator-endpoint", default="http://localhost:8001/v1",
|
||||
help="OpenAI-compatible vLLM base URL serving the orchestrator.")
|
||||
p.add_argument("--orchestrator-model", default="qwen3-8b")
|
||||
p.add_argument("--orchestrator-api-key", default="EMPTY")
|
||||
# Provenance stamps: when set, every written record carries these fields so
|
||||
# the JSONL is self-identifying once pooled across model families (gemma vs
|
||||
# qwen). Default None -> no stamping (existing qwen runs unaffected).
|
||||
p.add_argument("--gen-model", default=None,
|
||||
help="Full HF id of the generating orchestrator, stamped as "
|
||||
"record['gen_model'] (e.g. google/gemma-4-26B-A4B-it).")
|
||||
p.add_argument("--orchestrator-label", default=None,
|
||||
help="Short label stamped as record['orchestrator_model'] "
|
||||
"(e.g. gemma-4-26b).")
|
||||
# Local OSS model endpoints; repeatable "model_id=base_url" (e.g.
|
||||
# "Qwen/Qwen3.5-9B=http://localhost:8001/v1"). Unmapped local models are
|
||||
# still listed but served unconfigured (base_url=None).
|
||||
p.add_argument("--local-endpoint", action="append", default=[],
|
||||
metavar="MODEL_ID=URL",
|
||||
help="Local model id -> vLLM base URL (repeatable).")
|
||||
p.add_argument("--max-tasks", type=int, default=None,
|
||||
help="Cap on tasks (default: all of load_sft_tasks()).")
|
||||
p.add_argument("--skip-task-ids-from", action="append", default=[],
|
||||
metavar="GLOB",
|
||||
help="Glob of prior data.jsonl files; skip task_ids already "
|
||||
"generated there so this run only does unseen prompts "
|
||||
"(resume). Repeatable; applied before sharding.")
|
||||
p.add_argument("--samples-per-task", type=int, default=8)
|
||||
p.add_argument("--max-keep-per-task", type=int, default=1)
|
||||
p.add_argument("--max-turns", type=int, default=8)
|
||||
p.add_argument("--temperature", type=float, default=0.7)
|
||||
# Orchestrator completion cap per turn. The library default (4096) intermittently
|
||||
# truncates the final answer mid-sentence on longer reasoning turns; bump it so
|
||||
# the trace ends cleanly. Raise further (e.g. 16384) if traces still cut off.
|
||||
p.add_argument("--max-tokens", type=int, default=8192,
|
||||
help="Per-turn completion cap for the orchestrator (default 8192).")
|
||||
# Number of tasks rolled out concurrently against vLLM. Each in-flight task
|
||||
# issues its samples sequentially, so ~concurrency requests hit the server at
|
||||
# once. ~30-50 is comfortable on 1xL40S (prefix caching + continuous batching).
|
||||
p.add_argument("--concurrency", type=int, default=32,
|
||||
help="Tasks rolled out in parallel (default 32).")
|
||||
# Balanced-smoke pull: cap//4 from each of GeneralThought + OpenThoughts
|
||||
# code/math/science instead of the GeneralThought-only fast cap path. Use for a
|
||||
# representative smoke; the real run omits --max-tasks for the full 8K balanced set.
|
||||
p.add_argument("--balanced", action=argparse.BooleanOptionalAction, default=True,
|
||||
help="Draw an EVEN cross-domain sample (GeneralThought + "
|
||||
"OpenThoughts code/math/science). Default ON — pass "
|
||||
"--no-balanced for the old GeneralThought-only skew.")
|
||||
# Data-parallel sharding: split the (deterministic, seed-42) task list across
|
||||
# N independent driver processes, each pointed at its own orchestrator vLLM
|
||||
# replica. Shard i takes tasks[i::N] (a strided slice, so every shard stays
|
||||
# domain-balanced). Each writes its own --out file; concatenate the shard
|
||||
# JSONLs afterward. Lets the GPU-bound orchestrator scale across idle GPUs.
|
||||
p.add_argument("--shard-index", type=int, default=0,
|
||||
help="This shard's index in [0, shard-count).")
|
||||
p.add_argument("--shard-count", type=int, default=1,
|
||||
help="Total number of shards (default 1 = no sharding).")
|
||||
# Throughput knob: stop sampling a task as soon as max-keep-per-task passing
|
||||
# trajectories are found, instead of always running all --samples-per-task.
|
||||
# ~3-4x faster when most tasks solve early, but keeps the *first* passers
|
||||
# rather than the *cheapest* of N (drops the cost-optimisation signal).
|
||||
p.add_argument("--stop-at-keep", action="store_true",
|
||||
help="Short-circuit a task once max-keep passing samples found.")
|
||||
# Anonymize model experts (opaque random labels, uniform description, no cost
|
||||
# line, shuffled order) so the policy can't route on a model's name/position/
|
||||
# cost. The anon->real map is saved per record (metrics.anon_map) for analysis.
|
||||
p.add_argument("--anonymize-experts", action="store_true",
|
||||
help="Hide expert identity (random labels + shuffle) when routing.")
|
||||
# By default we keep EVERY rolled-out trajectory (correct + incorrect), each
|
||||
# tagged with ``correct`` (verifier verdict) and ``kept`` (the cheapest-correct
|
||||
# sample the rejection sampler would pick). Lets you compute accuracy / inspect
|
||||
# failures from the JSONL directly. --rejection-only restores the old
|
||||
# drop-the-failures behaviour (only cheapest-correct written).
|
||||
p.add_argument("--rejection-only", action="store_true",
|
||||
help="Drop incorrect rollouts; write only the cheapest-correct "
|
||||
"sample per task (the original behaviour).")
|
||||
# Rollouts call shell_exec / file_write with model-chosen relative paths
|
||||
# (e.g. ``solution.py``), which otherwise land in the repo root. Run them
|
||||
# from a throwaway scratch dir so generated files never dirty the tree.
|
||||
# gitignored via the existing ``scratch/`` rule.
|
||||
p.add_argument("--scratch-dir", default="scratch/sft-rollouts",
|
||||
help="CWD for rollouts; stray tool-written files go here.")
|
||||
args = p.parse_args(argv)
|
||||
|
||||
logging.basicConfig(level=logging.INFO, format="%(message)s")
|
||||
# Quiet the per-request HTTP / dataset-stream spam so the run log stays
|
||||
# readable (rollout calls and dataset shards otherwise flood it).
|
||||
for _noisy in ("httpx", "httpcore", "urllib3", "datasets", "fsspec",
|
||||
"huggingface_hub", "openai"):
|
||||
logging.getLogger(_noisy).setLevel(logging.WARNING)
|
||||
|
||||
# Every run lands in its OWN datetimed folder, all collected flat under
|
||||
# data/runs/ — so runs sort by time, never collide, and never litter data/.
|
||||
# --out is treated as an optional LABEL for the folder, not a path; the data
|
||||
# file inside is always data.jsonl (+ .stats.json/.pretty.json/.txt/.html).
|
||||
# Folder name: <MM-DD-HHMMam/pm>_<label> (local PT) e.g. 06-24-0349pm_gemma_20_anon
|
||||
stamp = time.strftime("%m-%d-%I%M%p").lower()
|
||||
label = Path(args.out).stem if args.out else "orchestrator_sft"
|
||||
run_dir = (Path("data/runs") / f"{stamp}_{label}").resolve()
|
||||
if run_dir.exists(): # parallel runs, same label+minute -> disambiguate
|
||||
run_dir = run_dir.with_name(f"{run_dir.name}-{os.getpid()}")
|
||||
out_p = run_dir / "data.jsonl"
|
||||
lock_p = out_p.with_suffix(out_p.suffix + ".lock")
|
||||
out_p.parent.mkdir(parents=True, exist_ok=True)
|
||||
lock_p.write_text(f"{stamp} pid={os.getpid()}")
|
||||
args.out = str(out_p)
|
||||
import atexit
|
||||
atexit.register(lambda: lock_p.exists() and lock_p.unlink())
|
||||
logging.info("Run dir: %s", run_dir)
|
||||
|
||||
# Enrich Braintrust rollout traces with run-level provenance (no-op-safe;
|
||||
# tracing.run_context() reads these). setdefault so an explicit env wins.
|
||||
os.environ.setdefault("OJ_RUN_LABEL", label)
|
||||
os.environ.setdefault("OJ_GEN_MODEL", args.gen_model or args.orchestrator_model)
|
||||
os.environ.setdefault("OJ_RUN_STAGE",
|
||||
"smoke" if args.max_tasks else "prod")
|
||||
os.environ["OJ_CFG_TEMPERATURE"] = str(args.temperature)
|
||||
os.environ["OJ_CFG_MAX_TURNS"] = str(args.max_turns)
|
||||
os.environ["OJ_CFG_ANONYMIZE"] = str(bool(args.anonymize_experts))
|
||||
os.environ["OJ_CFG_REJECTION_ONLY"] = str(bool(args.rejection_only))
|
||||
|
||||
# Sandbox the rollouts: chdir into a scratch dir so any file a rollout writes
|
||||
# (shell_exec redirects, file_write with a relative path) lands there instead
|
||||
# of the repo root. (--out is already absolute via run_dir.resolve() above.)
|
||||
scratch = Path(args.scratch_dir).resolve()
|
||||
scratch.mkdir(parents=True, exist_ok=True)
|
||||
os.chdir(scratch)
|
||||
logging.info("Rollout scratch dir (cwd): %s", scratch)
|
||||
|
||||
local_endpoints = {}
|
||||
for item in args.local_endpoint:
|
||||
model_id, _, url = item.partition("=")
|
||||
if model_id and url:
|
||||
local_endpoints[model_id] = url
|
||||
tools = orchestrator_catalog(local_endpoints=local_endpoints or None)
|
||||
# Drop the no-op ``think`` scratchpad tool: it adds turns/cost and is a
|
||||
# harness-leakage vector (the model narrates the routing rule inside a think
|
||||
# call, which the reasoning-scrub can't reach). The model still reasons in
|
||||
# its own <think> blocks — it just can't emit reasoning AS a tool call.
|
||||
tools = [t for t in tools if t.name != "think"]
|
||||
logging.info("Tool catalog (%d): %s", len(tools), [t.name for t in tools])
|
||||
|
||||
call_orch = make_call_orchestrator(
|
||||
args.orchestrator_model,
|
||||
base_url=args.orchestrator_endpoint,
|
||||
api_key=args.orchestrator_api_key,
|
||||
temperature=args.temperature,
|
||||
max_tokens=args.max_tokens,
|
||||
)
|
||||
dispatch = make_dispatch({})
|
||||
|
||||
# Build the provenance stamp once (None if neither flag given).
|
||||
record_extra = {}
|
||||
if args.gen_model:
|
||||
record_extra["gen_model"] = args.gen_model
|
||||
if args.orchestrator_label:
|
||||
record_extra["orchestrator_model"] = args.orchestrator_label
|
||||
record_extra = record_extra or None
|
||||
|
||||
def rollout_fn(task):
|
||||
try:
|
||||
return run_unified_rollout(
|
||||
task.instruction, tools,
|
||||
call_orchestrator=call_orch, dispatch=dispatch,
|
||||
max_turns=args.max_turns,
|
||||
anonymize=args.anonymize_experts,
|
||||
)
|
||||
except Exception as exc: # network/key failures shouldn't kill the run
|
||||
logging.warning("rollout failed for %s: %s", task.task_id, exc)
|
||||
return None
|
||||
|
||||
# cap= caps the task count for smoke runs; --balanced makes that small sample
|
||||
# cross-domain (GeneralThought + OpenThoughts code/math/science) instead of the
|
||||
# GeneralThought-only fast path. The real run omits --max-tasks (full 8K balanced).
|
||||
tasks = load_sft_tasks(cap=args.max_tasks, balanced=args.balanced)
|
||||
# Resume on unseen prompts: drop task_ids already generated in prior runs
|
||||
# (per-track seen set via --skip-task-ids-from). Filter BEFORE sharding so the
|
||||
# remaining unseen tasks stride evenly and never re-cover finished prompts.
|
||||
if args.skip_task_ids_from:
|
||||
import glob as _glob
|
||||
skip_ids = set()
|
||||
for pattern in args.skip_task_ids_from:
|
||||
for fp in _glob.glob(pattern):
|
||||
try:
|
||||
with open(fp) as fh:
|
||||
for line in fh:
|
||||
line = line.strip()
|
||||
if not line:
|
||||
continue
|
||||
try:
|
||||
tid = json.loads(line).get("task_id")
|
||||
except Exception:
|
||||
continue
|
||||
if tid:
|
||||
skip_ids.add(tid)
|
||||
except OSError:
|
||||
continue
|
||||
if skip_ids:
|
||||
n_before = len(tasks)
|
||||
tasks = [t for t in tasks if t.task_id not in skip_ids]
|
||||
logging.info("Resume: %d done task_ids -> skipping; %d of %d tasks remain",
|
||||
len(skip_ids), len(tasks), n_before)
|
||||
if args.shard_count > 1:
|
||||
n_all = len(tasks)
|
||||
tasks = tasks[args.shard_index::args.shard_count]
|
||||
logging.info("Shard %d/%d -> %d of %d tasks",
|
||||
args.shard_index, args.shard_count, len(tasks), n_all)
|
||||
from collections import Counter as _Counter
|
||||
_dom = _Counter(getattr(t, "domain", "unknown") for t in tasks)
|
||||
logging.info("Loaded %d SFT tasks | balanced=%s | domain split: %s",
|
||||
len(tasks), args.balanced,
|
||||
", ".join(f"{d}={n}" for d, n in sorted(_dom.items())))
|
||||
|
||||
stats = generate_sft_dataset(
|
||||
args.out,
|
||||
tasks=tasks,
|
||||
tools=tools,
|
||||
rollout_fn=rollout_fn,
|
||||
verify_fn=make_verifier(),
|
||||
samples_per_task=args.samples_per_task,
|
||||
max_keep_per_task=args.max_keep_per_task,
|
||||
reward_fn=lambda r: -r.cost_usd, # cheapest-correct gets highest reward
|
||||
concurrency=args.concurrency,
|
||||
stop_at_keep=args.stop_at_keep,
|
||||
keep_all=not args.rejection_only,
|
||||
record_extra=record_extra,
|
||||
)
|
||||
# Emit human-viewable companions (<out>.pretty.json + <out>.txt) so the
|
||||
# traces / full message turns can be inspected without un-escaping the JSONL.
|
||||
try:
|
||||
from render_sft_data import render # same scripts/orchestrator dir
|
||||
except ImportError:
|
||||
sys.path.insert(0, os.path.dirname(os.path.abspath(__file__)))
|
||||
from render_sft_data import render
|
||||
if stats.get("records_written", 0) > 0:
|
||||
rinfo = render(args.out)
|
||||
logging.info("Rendered viewable traces: %s + %s",
|
||||
rinfo["pretty"], rinfo["txt"])
|
||||
|
||||
print(json.dumps(stats, indent=2))
|
||||
return 0
|
||||
|
||||
|
||||
if __name__ == "__main__":
|
||||
raise SystemExit(main())
|
||||
@@ -0,0 +1,109 @@
|
||||
#!/usr/bin/env python
|
||||
"""Interactive REPL to chat with a trained orchestrator checkpoint.
|
||||
|
||||
Reuses the SAME routing loop the eval uses (OrchestratorBackend.generate_full),
|
||||
so what you see here is exactly how it behaves on the benchmarks: it reads your
|
||||
question, decides whether to answer itself / call a tool / delegate to an expert
|
||||
model, runs that rollout, and prints the final answer + a routing summary.
|
||||
|
||||
Prereqs:
|
||||
1. Serve the checkpoint you want on a vLLM port (see --endpoint). e.g. 8k:
|
||||
PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True CUDA_VISIBLE_DEVICES=0 \
|
||||
.venv/bin/python -m vllm.entrypoints.openai.api_server \
|
||||
--model "$OJ_WORK/qwen_eval_ckpts/8k_served" \
|
||||
--served-model-name sft-qwen-8k --port 8020 \
|
||||
--gpu-memory-utilization 0.88 --trust-remote-code --max-model-len 32768 \
|
||||
--enforce-eager --enable-auto-tool-choice --tool-call-parser qwen3_xml
|
||||
2. source .env (cloud experts GPT-5.5 / Opus need the API keys)
|
||||
|
||||
Usage:
|
||||
.venv/bin/python scripts/orchestrator/chat_orchestrator.py \
|
||||
--endpoint http://localhost:8020/v1 --model sft-qwen-8k
|
||||
|
||||
# point local expert tiers at their own vLLMs if you have them up (optional):
|
||||
# --local-endpoint Qwen/Qwen3.6-27B=http://localhost:8002/v1
|
||||
"""
|
||||
import argparse
|
||||
import sys
|
||||
import time
|
||||
|
||||
|
||||
def _parse_args(argv=None):
|
||||
p = argparse.ArgumentParser(description=__doc__,
|
||||
formatter_class=argparse.RawDescriptionHelpFormatter)
|
||||
p.add_argument("--endpoint", default="http://localhost:8020/v1",
|
||||
help="vLLM endpoint serving the orchestrator checkpoint.")
|
||||
p.add_argument("--model", default="sft-qwen-8k",
|
||||
help="Served-model-name of the checkpoint.")
|
||||
p.add_argument("--api-key", default="EMPTY", help="API key for the endpoint (local vLLM = EMPTY).")
|
||||
p.add_argument("--temperature", type=float, default=0.0)
|
||||
p.add_argument("--max-turns", type=int, default=8, help="Same default as the eval.")
|
||||
p.add_argument("--max-tokens", type=int, default=3000,
|
||||
help="Keep high: the model reasons a lot before it emits the delegation.")
|
||||
p.add_argument("--local-endpoint", action="append", default=[],
|
||||
help="MODEL_ID=URL for a local expert tier. Repeatable.")
|
||||
return p.parse_args(argv)
|
||||
|
||||
|
||||
def main(argv=None):
|
||||
args = _parse_args(argv)
|
||||
|
||||
local_endpoints = {}
|
||||
for pair in args.local_endpoint:
|
||||
if "=" not in pair:
|
||||
raise SystemExit(f"--local-endpoint expects MODEL_ID=URL, got: {pair!r}")
|
||||
mid, url = pair.split("=", 1)
|
||||
local_endpoints[mid.strip()] = url.strip()
|
||||
|
||||
from openjarvis.learning.intelligence.orchestrator.eval_backend import (
|
||||
OrchestratorBackend,
|
||||
)
|
||||
|
||||
backend = OrchestratorBackend(
|
||||
orchestrator_endpoint=args.endpoint,
|
||||
orchestrator_model=args.model,
|
||||
api_key=args.api_key,
|
||||
local_endpoints=local_endpoints,
|
||||
max_turns=args.max_turns,
|
||||
temperature=args.temperature,
|
||||
)
|
||||
|
||||
print(f"orchestrator: {args.model} @ {args.endpoint} "
|
||||
f"(temp={args.temperature}, max_turns={args.max_turns})")
|
||||
print("type a question and hit enter. Ctrl-D or 'quit' to exit.\n")
|
||||
|
||||
try:
|
||||
while True:
|
||||
try:
|
||||
prompt = input("you> ").strip()
|
||||
except EOFError:
|
||||
print()
|
||||
break
|
||||
if not prompt:
|
||||
continue
|
||||
if prompt.lower() in {"quit", "exit"}:
|
||||
break
|
||||
|
||||
started = time.time()
|
||||
full = backend.generate_full(
|
||||
prompt, model=args.model, temperature=args.temperature,
|
||||
max_tokens=args.max_tokens,
|
||||
)
|
||||
elapsed = time.time() - started
|
||||
|
||||
if full.get("error"):
|
||||
print(f"\n[ERROR] {full['error']}\n", file=sys.stderr)
|
||||
continue
|
||||
|
||||
print(f"\norch> {full.get('content', '').strip()}\n")
|
||||
print(f" [turns={full.get('turn_count', '?')} "
|
||||
f"tool_calls={full.get('tool_calls', '?')} "
|
||||
f"cost=${full.get('cost_usd', 0.0):.4f} "
|
||||
f"{elapsed:.1f}s]\n")
|
||||
finally:
|
||||
backend.close()
|
||||
return 0
|
||||
|
||||
|
||||
if __name__ == "__main__":
|
||||
raise SystemExit(main())
|
||||
@@ -0,0 +1,345 @@
|
||||
#!/usr/bin/env python
|
||||
"""Run the ToolOrchestra orchestrator over N-sample subsets of several
|
||||
benchmarks and score them with the existing eval infra.
|
||||
|
||||
Reuses, verbatim:
|
||||
* ``openjarvis.evals.cli._build_dataset`` / ``_build_scorer`` /
|
||||
``_build_judge_backend`` (loaders + scorers + judge wiring),
|
||||
* ``openjarvis.evals.core.runner.EvalRunner`` (parallel run + scoring),
|
||||
* ``openjarvis.evals.core.types.RunConfig`` (run config),
|
||||
|
||||
and plugs in our orchestrator as the "model" via
|
||||
``openjarvis.learning.intelligence.orchestrator.eval_backend.OrchestratorBackend``.
|
||||
|
||||
For each benchmark it runs ``EvalRunner(config, dataset, backend, scorer).run()``,
|
||||
collects accuracy / cost / latency, writes a combined ``summary.json``, and
|
||||
prints a table.
|
||||
|
||||
NOTE on the judge: the OpenAI key is dead, so the default judge model is a
|
||||
Gemini model (``gemini-2.5-flash``). The judge backend is still built via the
|
||||
``cloud`` engine (``_build_judge_backend``) — the model string selects the
|
||||
provider/route. Override with ``--judge-model`` / ``--judge-engine``.
|
||||
|
||||
Example
|
||||
-------
|
||||
.venv/bin/python scripts/orchestrator/eval_orchestrator.py \\
|
||||
--benchmarks gaia,mmlu_pro --n 100 \\
|
||||
--orchestrator-endpoint http://localhost:8001/v1 \\
|
||||
--orchestrator-model qwen3-8b \\
|
||||
--output-dir ~/.openjarvis/experiments/hybrid/runs/orch-eval
|
||||
"""
|
||||
|
||||
from __future__ import annotations
|
||||
|
||||
import argparse
|
||||
import json
|
||||
import sys
|
||||
import time
|
||||
from pathlib import Path
|
||||
from typing import Any, Dict, List, Optional
|
||||
|
||||
# Friendly aliases (underscored / shorthand) -> the EXACT registry keys used by
|
||||
# openjarvis.evals.cli.BENCHMARKS / _build_dataset / _build_scorer.
|
||||
BENCHMARK_ALIASES: Dict[str, str] = {
|
||||
"terminalbench_v2_1": "terminalbench-v2.1",
|
||||
"terminalbench-v2_1": "terminalbench-v2.1",
|
||||
"terminalbench_v21": "terminalbench-v2.1",
|
||||
"terminalbench-v2.1": "terminalbench-v2.1",
|
||||
"mmlu_pro": "mmlu-pro",
|
||||
"mmlu-pro": "mmlu-pro",
|
||||
"gaia": "gaia",
|
||||
"taubench": "taubench",
|
||||
"supergpqa": "supergpqa",
|
||||
}
|
||||
|
||||
DEFAULT_BENCHMARKS = "gaia,terminalbench_v2_1,taubench,mmlu_pro,supergpqa"
|
||||
|
||||
|
||||
def _normalize_benchmark(name: str) -> str:
|
||||
key = name.strip()
|
||||
return BENCHMARK_ALIASES.get(key, BENCHMARK_ALIASES.get(key.lower(), key))
|
||||
|
||||
|
||||
def _parse_args(argv: Optional[List[str]] = None) -> argparse.Namespace:
|
||||
p = argparse.ArgumentParser(
|
||||
description="Run the orchestrator over benchmark subsets and score them.",
|
||||
formatter_class=argparse.ArgumentDefaultsHelpFormatter,
|
||||
)
|
||||
p.add_argument(
|
||||
"--benchmarks",
|
||||
default=DEFAULT_BENCHMARKS,
|
||||
help="Comma-separated benchmark names (aliases normalized to registry keys).",
|
||||
)
|
||||
p.add_argument("--n", type=int, default=100, help="Samples per benchmark.")
|
||||
p.add_argument("--seed", type=int, default=42, help="Subset seed.")
|
||||
p.add_argument(
|
||||
"--orchestrator-endpoint",
|
||||
default="http://localhost:8001/v1",
|
||||
help="OpenAI-compatible base URL for the served orchestrator.",
|
||||
)
|
||||
p.add_argument(
|
||||
"--orchestrator-model",
|
||||
default="qwen3-8b",
|
||||
help="Served orchestrator model id.",
|
||||
)
|
||||
p.add_argument(
|
||||
"--orchestrator-api-key",
|
||||
default="EMPTY",
|
||||
help="API key for the orchestrator endpoint (EMPTY for local vLLM).",
|
||||
)
|
||||
p.add_argument(
|
||||
"--local-endpoint",
|
||||
action="append",
|
||||
default=[],
|
||||
metavar="MODEL_ID=URL",
|
||||
help=(
|
||||
"Wire a local OSS tool: map a catalog model id to its vLLM base "
|
||||
"URL, e.g. 'Qwen/Qwen3.5-9B=http://localhost:8003/v1'. Repeatable."
|
||||
),
|
||||
)
|
||||
p.add_argument(
|
||||
"--max-turns",
|
||||
type=int,
|
||||
default=8,
|
||||
help="Max orchestrator turns per sample.",
|
||||
)
|
||||
p.add_argument(
|
||||
"--temperature",
|
||||
type=float,
|
||||
default=1.0,
|
||||
help="Orchestrator sampling temperature.",
|
||||
)
|
||||
p.add_argument(
|
||||
"--max-workers",
|
||||
type=int,
|
||||
default=4,
|
||||
help="Parallel samples per benchmark.",
|
||||
)
|
||||
p.add_argument(
|
||||
"--output-dir",
|
||||
default="results/orchestrator-eval",
|
||||
help="Directory for per-benchmark JSONL + combined summary.json.",
|
||||
)
|
||||
# NOTE: OpenAI key is dead — default the judge to a Gemini model.
|
||||
p.add_argument(
|
||||
"--judge-model",
|
||||
default="gemini-2.5-flash",
|
||||
help="LLM-judge model (default Gemini — OpenAI key is dead).",
|
||||
)
|
||||
p.add_argument(
|
||||
"--judge-engine",
|
||||
default="cloud",
|
||||
help="Engine key for the judge backend.",
|
||||
)
|
||||
return p.parse_args(argv)
|
||||
|
||||
|
||||
def _run_one(
|
||||
benchmark: str,
|
||||
*,
|
||||
n: int,
|
||||
seed: int,
|
||||
backend,
|
||||
judge_model: str,
|
||||
judge_engine: str,
|
||||
max_workers: int,
|
||||
orchestrator_model: str,
|
||||
output_dir: Path,
|
||||
) -> Dict[str, Any]:
|
||||
"""Run + score one benchmark, returning a row dict (or an error row)."""
|
||||
# Imports here so --help works without importing the (heavy) eval stack.
|
||||
from openjarvis.evals.cli import (
|
||||
_build_dataset,
|
||||
_build_judge_backend,
|
||||
_build_scorer,
|
||||
)
|
||||
from openjarvis.evals.core.runner import EvalRunner
|
||||
from openjarvis.evals.core.types import RunConfig
|
||||
|
||||
output_path = output_dir / f"{benchmark.replace('.', '_')}_orchestrator.jsonl"
|
||||
|
||||
dataset = _build_dataset(benchmark)
|
||||
# The runner reloads internally; we also load here per spec so callers can
|
||||
# inspect/size the subset up front. load() is idempotent for these sets.
|
||||
dataset.load(max_samples=n, seed=seed)
|
||||
|
||||
judge_backend = _build_judge_backend(judge_model, engine_key=judge_engine)
|
||||
scorer = _build_scorer(benchmark, judge_backend, judge_model)
|
||||
|
||||
config = RunConfig(
|
||||
benchmark=benchmark,
|
||||
backend="orchestrator",
|
||||
model=orchestrator_model,
|
||||
max_samples=n,
|
||||
max_workers=max_workers,
|
||||
seed=seed,
|
||||
judge_model=judge_model,
|
||||
judge_engine=judge_engine,
|
||||
output_path=str(output_path),
|
||||
)
|
||||
|
||||
runner = EvalRunner(config, dataset, backend, scorer)
|
||||
started = time.time()
|
||||
try:
|
||||
summary = runner.run()
|
||||
finally:
|
||||
if judge_backend is not None:
|
||||
judge_backend.close()
|
||||
elapsed = time.time() - started
|
||||
|
||||
return {
|
||||
"benchmark": benchmark,
|
||||
"accuracy": summary.accuracy,
|
||||
"scored_samples": summary.scored_samples,
|
||||
"correct": summary.correct,
|
||||
"errors": summary.errors,
|
||||
"total_samples": summary.total_samples,
|
||||
"mean_latency_seconds": summary.mean_latency_seconds,
|
||||
"total_cost_usd": summary.total_cost_usd,
|
||||
"mean_continuous_score": summary.mean_continuous_score,
|
||||
"wall_seconds": round(elapsed, 2),
|
||||
"output_path": str(output_path),
|
||||
}
|
||||
|
||||
|
||||
def _print_table(rows: List[Dict[str, Any]]) -> None:
|
||||
cols = [
|
||||
("benchmark", 22, "{}"),
|
||||
("accuracy", 9, "{:.4f}"),
|
||||
("correct", 8, "{}"),
|
||||
("scored", 7, "{}"),
|
||||
("errors", 7, "{}"),
|
||||
("cost($)", 10, "{:.4f}"),
|
||||
("lat(s)", 9, "{:.2f}"),
|
||||
("wall(s)", 9, "{:.1f}"),
|
||||
]
|
||||
header = " ".join(f"{name:<{w}}" for name, w, _ in cols)
|
||||
print("\n" + header)
|
||||
print("-" * len(header))
|
||||
for r in rows:
|
||||
if r.get("error"):
|
||||
print(f"{r['benchmark']:<22} ERROR: {r['error']}")
|
||||
continue
|
||||
vals = {
|
||||
"benchmark": r["benchmark"],
|
||||
"accuracy": r["accuracy"],
|
||||
"correct": r["correct"],
|
||||
"scored": r["scored_samples"],
|
||||
"errors": r["errors"],
|
||||
"cost($)": r["total_cost_usd"],
|
||||
"lat(s)": r["mean_latency_seconds"],
|
||||
"wall(s)": r["wall_seconds"],
|
||||
}
|
||||
cells = []
|
||||
for name, w, fmt in cols:
|
||||
v = vals[name]
|
||||
try:
|
||||
s = fmt.format(v)
|
||||
except (ValueError, TypeError):
|
||||
s = str(v)
|
||||
cells.append(f"{s:<{w}}")
|
||||
print(" ".join(cells))
|
||||
|
||||
|
||||
def main(argv: Optional[List[str]] = None) -> int:
|
||||
args = _parse_args(argv)
|
||||
|
||||
benchmarks = [
|
||||
_normalize_benchmark(b) for b in args.benchmarks.split(",") if b.strip()
|
||||
]
|
||||
output_dir = Path(args.output_dir).expanduser()
|
||||
output_dir.mkdir(parents=True, exist_ok=True)
|
||||
|
||||
# Build the orchestrator backend once (stateless across benchmarks).
|
||||
from openjarvis.learning.intelligence.orchestrator.eval_backend import (
|
||||
OrchestratorBackend,
|
||||
)
|
||||
|
||||
local_endpoints: Dict[str, str] = {}
|
||||
for pair in args.local_endpoint:
|
||||
if "=" not in pair:
|
||||
raise SystemExit(
|
||||
f"--local-endpoint expects MODEL_ID=URL, got: {pair!r}"
|
||||
)
|
||||
model_id, url = pair.split("=", 1)
|
||||
local_endpoints[model_id.strip()] = url.strip()
|
||||
|
||||
backend = OrchestratorBackend(
|
||||
orchestrator_endpoint=args.orchestrator_endpoint,
|
||||
orchestrator_model=args.orchestrator_model,
|
||||
api_key=args.orchestrator_api_key,
|
||||
local_endpoints=local_endpoints,
|
||||
max_turns=args.max_turns,
|
||||
temperature=args.temperature,
|
||||
)
|
||||
|
||||
rows: List[Dict[str, Any]] = []
|
||||
try:
|
||||
for benchmark in benchmarks:
|
||||
print(f"\n=== {benchmark} (n={args.n}, seed={args.seed}) ===")
|
||||
try:
|
||||
row = _run_one(
|
||||
benchmark,
|
||||
n=args.n,
|
||||
seed=args.seed,
|
||||
backend=backend,
|
||||
judge_model=args.judge_model,
|
||||
judge_engine=args.judge_engine,
|
||||
max_workers=args.max_workers,
|
||||
orchestrator_model=args.orchestrator_model,
|
||||
output_dir=output_dir,
|
||||
)
|
||||
print(
|
||||
f" accuracy={row['accuracy']:.4f} "
|
||||
f"({row['correct']}/{row['scored_samples']}) "
|
||||
f"errors={row['errors']} cost=${row['total_cost_usd']:.4f}"
|
||||
)
|
||||
except Exception as exc: # noqa: BLE001 - one bad bench shouldn't kill all
|
||||
print(f" FAILED: {type(exc).__name__}: {exc}", file=sys.stderr)
|
||||
row = {"benchmark": benchmark, "error": f"{type(exc).__name__}: {exc}"}
|
||||
rows.append(row)
|
||||
finally:
|
||||
backend.close()
|
||||
|
||||
# Resolve the EXACT model behind the served alias (e.g. "gemma-sft" ->
|
||||
# the real id + checkpoint path vLLM reports), so results are self-identifying
|
||||
# instead of an opaque label. Best-effort: never fail the summary over it.
|
||||
served_id, served_path = args.orchestrator_model, None
|
||||
try:
|
||||
import urllib.request
|
||||
req = urllib.request.Request(
|
||||
args.orchestrator_endpoint.rstrip("/") + "/models",
|
||||
headers={"Authorization": f"Bearer {args.orchestrator_api_key}"},
|
||||
)
|
||||
with urllib.request.urlopen(req, timeout=10) as resp:
|
||||
data = json.load(resp).get("data", [])
|
||||
if data:
|
||||
served_id = data[0].get("id", served_id)
|
||||
# vLLM reports the loaded checkpoint path in `root` (or `parent`).
|
||||
served_path = data[0].get("root") or data[0].get("parent")
|
||||
except Exception:
|
||||
pass
|
||||
|
||||
combined = {
|
||||
"orchestrator_model": args.orchestrator_model,
|
||||
"orchestrator_served_id": served_id,
|
||||
"orchestrator_model_path": served_path,
|
||||
"orchestrator_endpoint": args.orchestrator_endpoint,
|
||||
"judge_model": args.judge_model,
|
||||
"judge_engine": args.judge_engine,
|
||||
"n": args.n,
|
||||
"seed": args.seed,
|
||||
"max_turns": args.max_turns,
|
||||
"benchmarks": rows,
|
||||
}
|
||||
summary_path = output_dir / "summary.json"
|
||||
with open(summary_path, "w") as f:
|
||||
json.dump(combined, f, indent=2, default=str)
|
||||
|
||||
_print_table(rows)
|
||||
print(f"\nCombined summary written to {summary_path}")
|
||||
return 0
|
||||
|
||||
|
||||
if __name__ == "__main__":
|
||||
raise SystemExit(main())
|
||||
@@ -0,0 +1,239 @@
|
||||
#!/usr/bin/env python
|
||||
"""Render graded orchestrator-eval JSONL records into clean, readable .txt files.
|
||||
|
||||
Companion to ``format_sft_sample.py`` (same banner/indent visual style), but
|
||||
for the *scored* eval outputs written by ``eval_orchestrator.py``:
|
||||
|
||||
results/orch-eval-<model>-<tranche>-<suite>/<benchmark>_orchestrator.jsonl
|
||||
|
||||
Each record has: record_id, benchmark, model, model_answer (the model's
|
||||
FINAL_ANSWER string), is_correct, score, latency_seconds, cost_usd, error, and
|
||||
scoring_metadata. The full step-by-step routing trace is NOT saved — only the
|
||||
final answer + gold + score — so this renderer cannot show intermediate steps.
|
||||
|
||||
The QUESTION text is not in the result JSONL; it's loaded from the benchmark
|
||||
dataset by ``record_id`` (via ``openjarvis.evals.cli._build_dataset``). If the
|
||||
dataset can't be loaded / the id doesn't match, we render everything else and
|
||||
show "(question unavailable)".
|
||||
|
||||
Gold answer lives in ``scoring_metadata`` (shape varies by scorer):
|
||||
* MCQ (mmlu-pro, supergpqa): ``reference_letter``.
|
||||
* LLM-judge (gaia): parsed from ``raw_judge_output`` (`gold target "..."`),
|
||||
plus the judge's ``reasoning``.
|
||||
* anything else: dumped verbatim under SCORING (raw).
|
||||
|
||||
Usage:
|
||||
.venv/bin/python scripts/orchestrator/format_eval_sample.py \
|
||||
--input results/orch-eval-gemma-2k-full/gaia_orchestrator.jsonl --n 2
|
||||
|
||||
... --input <file> --lines 1,5,42 # specific 1-indexed lines
|
||||
... --input <file> --all # one .txt per record
|
||||
... --input <file> --all --only-wrong # only incorrect samples
|
||||
"""
|
||||
from __future__ import annotations
|
||||
|
||||
import argparse
|
||||
import json
|
||||
import re
|
||||
from pathlib import Path
|
||||
|
||||
WIDTH = 80
|
||||
|
||||
# Friendly aliases -> the registry keys _build_dataset expects (mirrors
|
||||
# eval_orchestrator.BENCHMARK_ALIASES for the ones that write .jsonl here).
|
||||
BENCHMARK_ALIASES = {
|
||||
"mmlu_pro": "mmlu-pro",
|
||||
"mmlu-pro": "mmlu-pro",
|
||||
"gaia": "gaia",
|
||||
"taubench": "taubench",
|
||||
"supergpqa": "supergpqa",
|
||||
"terminalbench_v2_1": "terminalbench-v2.1",
|
||||
"terminalbench-v2.1": "terminalbench-v2.1",
|
||||
}
|
||||
|
||||
|
||||
def _banner(label: str) -> str:
|
||||
"""Full-width headline: ``━━━ LABEL ━━━━━━…``."""
|
||||
prefix = f"━━━ {label} "
|
||||
fill = "━" * max(3, WIDTH - len(prefix))
|
||||
return f"\n{prefix}{fill}\n"
|
||||
|
||||
|
||||
def _indent(text: str, pad: str = " ") -> str:
|
||||
text = (text or "").rstrip()
|
||||
return "\n".join(pad + ln if ln.strip() else ln for ln in text.splitlines())
|
||||
|
||||
|
||||
def _normalize_benchmark(name: str) -> str:
|
||||
key = (name or "").strip()
|
||||
return BENCHMARK_ALIASES.get(key, BENCHMARK_ALIASES.get(key.lower(), key))
|
||||
|
||||
|
||||
# ---------------------------------------------------------------------------
|
||||
# Question map: record_id -> problem text, loaded from the benchmark dataset.
|
||||
# ---------------------------------------------------------------------------
|
||||
def build_question_map(benchmark: str, n: int, seed: int = 42) -> dict:
|
||||
"""Best-effort {record_id: problem}. Returns {} if the dataset can't load."""
|
||||
try:
|
||||
from openjarvis.evals.cli import _build_dataset
|
||||
|
||||
ds = _build_dataset(_normalize_benchmark(benchmark))
|
||||
ds.load(max_samples=n, seed=seed)
|
||||
return {r.record_id: r.problem for r in ds.iter_records()}
|
||||
except Exception as exc: # noqa: BLE001 - never fail the render over this
|
||||
print(f" [warn] could not load dataset for {benchmark!r}: "
|
||||
f"{type(exc).__name__}: {exc}")
|
||||
return {}
|
||||
|
||||
|
||||
# ---------------------------------------------------------------------------
|
||||
# Gold / judge parsing out of scoring_metadata.
|
||||
# ---------------------------------------------------------------------------
|
||||
_GOLD_TARGET_RE = re.compile(r'gold target\s*"([^"]*)"', re.IGNORECASE)
|
||||
_JUDGE_REASONING_RE = re.compile(
|
||||
r"reasoning:\s*(.*?)(?:\n\s*correct:|\Z)", re.IGNORECASE | re.DOTALL
|
||||
)
|
||||
|
||||
|
||||
def parse_scoring(meta: dict) -> dict:
|
||||
"""Return {gold, judge, kind, clean}. ``clean`` False => dump raw block.
|
||||
|
||||
kind is one of 'mcq' | 'judge' | 'unknown'.
|
||||
"""
|
||||
if not isinstance(meta, dict):
|
||||
return {"gold": None, "judge": None, "kind": "unknown", "clean": False}
|
||||
|
||||
# MCQ scorer (mmlu-pro / supergpqa).
|
||||
if "reference_letter" in meta:
|
||||
return {
|
||||
"gold": meta.get("reference_letter"),
|
||||
"judge": None,
|
||||
"kind": "mcq",
|
||||
"clean": True,
|
||||
}
|
||||
|
||||
# LLM-judge scorer (gaia).
|
||||
raw = meta.get("raw_judge_output")
|
||||
if raw:
|
||||
gold_m = _GOLD_TARGET_RE.search(raw)
|
||||
reason_m = _JUDGE_REASONING_RE.search(raw)
|
||||
return {
|
||||
"gold": gold_m.group(1) if gold_m else None,
|
||||
"judge": (reason_m.group(1).strip() if reason_m else raw.strip()),
|
||||
"kind": "judge",
|
||||
"clean": True,
|
||||
}
|
||||
|
||||
# Unknown shape (e.g. {"reason": "no_choice_letter_extracted"}).
|
||||
return {"gold": None, "judge": None, "kind": "unknown", "clean": False}
|
||||
|
||||
|
||||
# ---------------------------------------------------------------------------
|
||||
# Render one record.
|
||||
# ---------------------------------------------------------------------------
|
||||
def format_record(rec: dict, question: str | None) -> str:
|
||||
parsed = parse_scoring(rec.get("scoring_metadata"))
|
||||
is_correct = rec.get("is_correct")
|
||||
verdict = "CORRECT" if is_correct else ("INCORRECT" if is_correct is False else "UNSCORED")
|
||||
|
||||
def _fmt(v, fmt):
|
||||
try:
|
||||
return fmt.format(v)
|
||||
except (ValueError, TypeError):
|
||||
return str(v)
|
||||
|
||||
top = " · ".join([
|
||||
str(rec.get("record_id", "?")),
|
||||
str(rec.get("benchmark", "?")),
|
||||
verdict,
|
||||
f"score {_fmt(rec.get('score'), '{:.3f}')}",
|
||||
f"lat {_fmt(rec.get('latency_seconds'), '{:.1f}')}s",
|
||||
f"cost ${_fmt(rec.get('cost_usd'), '{:.4f}')}",
|
||||
])
|
||||
parts = [top]
|
||||
|
||||
err = rec.get("error")
|
||||
if err:
|
||||
parts.append(f" error: {err}")
|
||||
|
||||
parts.append(_banner("QUESTION"))
|
||||
parts.append(_indent(question if question else "(question unavailable)"))
|
||||
|
||||
parts.append(_banner("MODEL ANSWER"))
|
||||
parts.append(_indent(str(rec.get("model_answer") or "(empty)")))
|
||||
|
||||
parts.append(_banner("GOLD"))
|
||||
gold = parsed["gold"]
|
||||
parts.append(_indent(str(gold) if gold not in (None, "") else "(gold unavailable — see SCORING below)"))
|
||||
|
||||
if parsed["kind"] == "judge" and parsed["judge"]:
|
||||
parts.append(_banner("JUDGE"))
|
||||
parts.append(_indent(parsed["judge"]))
|
||||
|
||||
if not parsed["clean"]:
|
||||
parts.append(_banner("SCORING (raw)"))
|
||||
parts.append(_indent(json.dumps(rec.get("scoring_metadata"), indent=2, default=str)))
|
||||
|
||||
return "\n".join(parts).rstrip() + "\n"
|
||||
|
||||
|
||||
def main(argv=None):
|
||||
p = argparse.ArgumentParser(
|
||||
description=__doc__, formatter_class=argparse.RawDescriptionHelpFormatter
|
||||
)
|
||||
p.add_argument("--input", required=True, help="Graded *_orchestrator.jsonl file.")
|
||||
p.add_argument("--out-dir", default="results/formatted", help="Where the .txt files go.")
|
||||
p.add_argument("--seed", type=int, default=42, help="Subset seed used at eval time.")
|
||||
p.add_argument("--only-wrong", action="store_true",
|
||||
help="Render only incorrect/unscored samples.")
|
||||
g = p.add_mutually_exclusive_group()
|
||||
g.add_argument("--n", type=int, default=1, help="Format the first N records.")
|
||||
g.add_argument("--lines", help="Comma-separated 1-indexed line numbers.")
|
||||
g.add_argument("--all", action="store_true", help="Format every record.")
|
||||
args = p.parse_args(argv)
|
||||
|
||||
src = Path(args.input)
|
||||
out_dir = Path(args.out_dir)
|
||||
out_dir.mkdir(parents=True, exist_ok=True)
|
||||
|
||||
records = [json.loads(l) for l in src.open() if l.strip()]
|
||||
if not records:
|
||||
print("no records in input")
|
||||
return 1
|
||||
|
||||
# Pick indices first (so we only load the dataset once, sized to the file).
|
||||
if args.lines:
|
||||
idxs = [int(x) - 1 for x in args.lines.split(",") if x.strip()]
|
||||
elif args.all:
|
||||
idxs = list(range(len(records)))
|
||||
else:
|
||||
idxs = list(range(min(args.n, len(records))))
|
||||
idxs = [i for i in idxs if 0 <= i < len(records)]
|
||||
|
||||
if args.only_wrong:
|
||||
idxs = [i for i in idxs if not records[i].get("is_correct")]
|
||||
|
||||
benchmark = _normalize_benchmark(records[0].get("benchmark", src.stem.split("_")[0]))
|
||||
# Load the whole subset so any selected id resolves (the eval-time subset is
|
||||
# the first len(records) of seed=42).
|
||||
qmap = build_question_map(benchmark, n=len(records), seed=args.seed)
|
||||
|
||||
written = []
|
||||
for i in idxs:
|
||||
rec = records[i]
|
||||
rid = str(rec.get("record_id", i))
|
||||
bench = rec.get("benchmark", benchmark)
|
||||
fname = re.sub(r"[^A-Za-z0-9_.-]", "_", f"{bench}__{rid}")[:120] + ".txt"
|
||||
out = out_dir / fname
|
||||
out.write_text(format_record(rec, qmap.get(rid)))
|
||||
written.append(out)
|
||||
|
||||
for w in written:
|
||||
print(w)
|
||||
print(f"\nwrote {len(written)} file(s) to {out_dir}/"
|
||||
+ (" (dataset unavailable — questions omitted)" if not qmap else ""))
|
||||
return 0
|
||||
|
||||
|
||||
if __name__ == "__main__":
|
||||
raise SystemExit(main())
|
||||
@@ -0,0 +1,120 @@
|
||||
#!/usr/bin/env python
|
||||
"""Carve train / holdout / overfit100 splits from a clean orchestrator-SFT pool.
|
||||
|
||||
Replaces make_tranches.py + make_gemma_tranches.py, which carved 1k/2k/4k/8k
|
||||
data-scaling tranches from the June pool. That lineage is dead: the tranches it
|
||||
produced trained checkpoints that evaluated at base ~= 1k ~= 2k (the pool taught
|
||||
the orchestrator to re-derive answers itself instead of routing), and the pool
|
||||
was deleted. See data/DELETED_ERA1_MANIFEST.md.
|
||||
|
||||
Naming is uniform: ``{name}_orch_{split}_{date}.jsonl`` with name in
|
||||
{qwen, gemma, pooled}. Pass several pools to merge-and-restratify into `pooled`.
|
||||
|
||||
The holdout is domain-stratified so val-loss is leak-free, and `overfit100` is a
|
||||
strict prefix of train (it is a memorisation sanity check — the model must be
|
||||
able to fit 100 rows it *has* seen, else the format/masking is broken).
|
||||
Deterministic (seed 42).
|
||||
|
||||
# one pool -> qwen_orch_{train,holdout,overfit100}_0711.jsonl
|
||||
python scripts/orchestrator/make_splits.py --name qwen \
|
||||
--pool data/sft_variants/qwen_orch_clean_0711.jsonl
|
||||
|
||||
# merge both -> pooled_orch_{train,holdout,overfit100}_0711.jsonl
|
||||
python scripts/orchestrator/make_splits.py --name pooled \
|
||||
--pool data/sft_variants/qwen_orch_clean_0711.jsonl \
|
||||
--pool data/sft_variants/gemma_orch_clean_0711.jsonl
|
||||
"""
|
||||
import argparse
|
||||
import json
|
||||
import os
|
||||
import random
|
||||
import sys
|
||||
from collections import defaultdict
|
||||
from datetime import datetime
|
||||
from pathlib import Path
|
||||
|
||||
OUT = Path("data/sft_variants")
|
||||
SEED = 42
|
||||
OVERFIT_N = 100
|
||||
# Orchestrator that generated each pool — stamped onto every row so provenance
|
||||
# survives a merge (the filename encodes it too, but the field is cheap insurance
|
||||
# and is what the `pooled` splits rely on to stay attributable).
|
||||
ORCH_MODEL = {"qwen": "qwen3.5-9b", "gemma": "gemma-4-26b"}
|
||||
|
||||
|
||||
def main() -> int:
|
||||
ap = argparse.ArgumentParser(description=__doc__,
|
||||
formatter_class=argparse.RawDescriptionHelpFormatter)
|
||||
ap.add_argument("--name", required=True,
|
||||
help="split family: qwen | gemma | pooled (drives the filename)")
|
||||
ap.add_argument("--pool", action="append", required=True, metavar="PATH",
|
||||
help="clean pool jsonl; repeat to merge (use with --name pooled)")
|
||||
ap.add_argument("--holdout-frac", type=float, default=0.15,
|
||||
help="fraction held out, domain-stratified (default 0.15)")
|
||||
ap.add_argument("--date", default=datetime.now().strftime("%m%d"),
|
||||
help="date tag in the filename (default: today, MMDD)")
|
||||
ap.add_argument("--out", type=Path, default=OUT)
|
||||
args = ap.parse_args()
|
||||
|
||||
rows = []
|
||||
for p in args.pool:
|
||||
path = Path(p)
|
||||
if not path.exists():
|
||||
print(f"!! pool not found: {path}", file=sys.stderr)
|
||||
return 1
|
||||
# Infer the source orchestrator from the filename so merged pools stay
|
||||
# attributable; don't clobber a stamp the pool already carries.
|
||||
stem = path.name.split("_", 1)[0]
|
||||
loaded = [json.loads(line) for line in path.open() if line.strip()]
|
||||
for r in loaded:
|
||||
r.setdefault("orchestrator_model", ORCH_MODEL.get(stem, stem))
|
||||
rows.extend(loaded)
|
||||
print(f"loaded {len(loaded):>5} rows from {path}")
|
||||
print(f"pool total: {len(rows)}")
|
||||
|
||||
# Stratified holdout: per-domain, proportional to domain size.
|
||||
by_dom = defaultdict(list)
|
||||
for i, r in enumerate(rows):
|
||||
by_dom[r.get("domain", "misc")].append(i)
|
||||
rng = random.Random(SEED)
|
||||
holdout_idx: set = set()
|
||||
for dom, idxs in sorted(by_dom.items()):
|
||||
k = round(args.holdout_frac * len(idxs))
|
||||
pick = rng.sample(idxs, min(k, len(idxs)))
|
||||
holdout_idx.update(pick)
|
||||
print(f" {dom:<9} pool={len(idxs):>4} holdout={len(pick)}")
|
||||
|
||||
holdout = [rows[i] for i in sorted(holdout_idx)]
|
||||
train = [r for i, r in enumerate(rows) if i not in holdout_idx] # order preserved
|
||||
overfit = train[:OVERFIT_N]
|
||||
print(f"train={len(train)} holdout={len(holdout)} overfit={len(overfit)}")
|
||||
|
||||
args.out.mkdir(parents=True, exist_ok=True)
|
||||
written = {}
|
||||
for split, data in (("train", train), ("holdout", holdout),
|
||||
(f"overfit{OVERFIT_N}", overfit)):
|
||||
f = args.out / f"{args.name}_orch_{split}_{args.date}.jsonl"
|
||||
f.write_text("".join(json.dumps(r) + "\n" for r in data))
|
||||
written[split] = f
|
||||
print(f"wrote {f} ({len(data)})")
|
||||
|
||||
# ---- auto-upload train + holdout to Braintrust ----
|
||||
# Gated by OJ_BRAINTRUST_AUTOUPLOAD (default ON); a total no-op / never raises
|
||||
# if the key/pkg is missing or the upload errors (see upload_to_braintrust).
|
||||
sys.path.insert(0, str(Path(__file__).resolve().parent))
|
||||
try:
|
||||
from upload_to_braintrust import autoupload
|
||||
|
||||
autoupload(
|
||||
[f"{written['train']}={args.name}_train_{args.date}",
|
||||
f"{written['holdout']}={args.name}_holdout_{args.date}"],
|
||||
run_label=os.getenv("OJ_RUN_LABEL", f"{args.name}_splits_{args.date}"),
|
||||
description=f"{args.name} orchestrator-SFT splits (make_splits.py)",
|
||||
)
|
||||
except Exception as exc: # telemetry must never break the data pipeline
|
||||
print(f"[braintrust] autoupload hook skipped ({exc})")
|
||||
return 0
|
||||
|
||||
|
||||
if __name__ == "__main__":
|
||||
raise SystemExit(main())
|
||||
@@ -0,0 +1,290 @@
|
||||
#!/usr/bin/env python
|
||||
"""Render an orchestrator SFT JSONL into human-viewable companions.
|
||||
|
||||
The trainer-facing ``*.jsonl`` packs every trajectory onto one line with all
|
||||
newlines escaped, which is unreadable by eye. This writes two siblings next to
|
||||
it so you can actually inspect what the orchestrator did:
|
||||
|
||||
* ``<name>.pretty.json`` — the same records as an indented JSON array.
|
||||
* ``<name>.txt`` — a clean plain-text transcript. The shared system
|
||||
prompt + tool catalog (identical across every record) is printed ONCE at the
|
||||
top; each trajectory then shows just its turns: the user question, the
|
||||
assistant's reasoning/answer, tool calls rendered as ``-> name({args})``, and
|
||||
tool results truncated so a single web-search dump doesn't bury the trace.
|
||||
|
||||
Usage:
|
||||
.venv/bin/python scripts/orchestrator/render_sft_data.py <path.jsonl>
|
||||
"""
|
||||
|
||||
from __future__ import annotations
|
||||
|
||||
import html as _htmllib
|
||||
import json
|
||||
import re
|
||||
import sys
|
||||
from pathlib import Path
|
||||
from typing import List
|
||||
|
||||
# Tool observations (web_search dumps, code stdout, file reads) can be huge; cap
|
||||
# them in the transcript so the actual orchestration stays legible. The full
|
||||
# content is always in the .jsonl / .pretty.json.
|
||||
_OBS_CAP = 1200
|
||||
_RULE = "-" * 80
|
||||
_TOOL_CALL_RE = re.compile(r"<tool_call>\s*(\{.*?\})\s*</tool_call>", re.DOTALL)
|
||||
|
||||
|
||||
def _parse_tool_calls(content: str):
|
||||
"""Parse the model's ``<tool_call>{json}</tool_call>`` tags into a list of
|
||||
parsed ``[(name, arguments_dict), ...]`` calls, de-duplicating the
|
||||
double-emit the template produces (each call shows up twice)."""
|
||||
seen = []
|
||||
for m in _TOOL_CALL_RE.finditer(content):
|
||||
try:
|
||||
call = json.loads(m.group(1))
|
||||
item = (call.get("name"), call.get("arguments", {}))
|
||||
except Exception:
|
||||
item = (None, {"_raw": m.group(1)})
|
||||
key = json.dumps(item, ensure_ascii=False, sort_keys=True, default=str)
|
||||
if not seen or seen[-1][0] != key: # collapse the adjacent duplicate
|
||||
seen.append((key, item))
|
||||
return [it for _, it in seen]
|
||||
|
||||
|
||||
def _call_label(name, anon_map) -> str:
|
||||
"""Display a call name; if it's an anonymized expert label, append the real
|
||||
model it maps to (from the record's ``metrics.anon_map``)."""
|
||||
real = (anon_map or {}).get(name)
|
||||
return f"{name} -> {real}" if real else f"{name}"
|
||||
|
||||
|
||||
def _fmt_tool_calls(content: str, anon_map=None) -> str:
|
||||
"""Plain-text tool calls with real (un-escaped) newlines in each arg value."""
|
||||
blocks = []
|
||||
for name, args in _parse_tool_calls(content):
|
||||
lines = [f"-> {_call_label(name, anon_map)}"]
|
||||
if isinstance(args, dict):
|
||||
for k, v in args.items():
|
||||
val = v if isinstance(v, str) else json.dumps(v, ensure_ascii=False, indent=2)
|
||||
indented = "\n".join(" " + ln for ln in val.splitlines())
|
||||
lines.append(f" {k}:\n{indented}")
|
||||
blocks.append("\n".join(lines))
|
||||
return "\n".join(blocks)
|
||||
|
||||
|
||||
def _clip(text: str, cap: int) -> str:
|
||||
text = text or ""
|
||||
if len(text) <= cap:
|
||||
return text
|
||||
return f"{text[:cap]}\n ... [truncated {len(text) - cap} chars]"
|
||||
|
||||
|
||||
def _render_turn(turn: dict, anon_map=None) -> str:
|
||||
role = str(turn.get("role", "?")).lower()
|
||||
content = turn.get("content", "") or ""
|
||||
|
||||
if role == "user":
|
||||
return f"USER:\n{content.strip()}"
|
||||
|
||||
if role in ("tool", "function"):
|
||||
name = turn.get("name", "tool")
|
||||
return f"TOOL [{name}]:\n{_clip(content.strip(), _OBS_CAP)}"
|
||||
|
||||
if role == "assistant":
|
||||
calls = _fmt_tool_calls(content, anon_map)
|
||||
# Strip the raw <tool_call> tags out of the prose so it isn't shown twice.
|
||||
prose = _TOOL_CALL_RE.sub("", content).strip()
|
||||
parts = []
|
||||
if prose:
|
||||
parts.append(f"ASSISTANT:\n{prose}")
|
||||
if calls:
|
||||
parts.append(f"ASSISTANT calls:\n{calls}")
|
||||
return "\n".join(parts) if parts else "ASSISTANT: (empty)"
|
||||
|
||||
# system handled separately (printed once); fall through for anything else
|
||||
return f"{role.upper()}:\n{content.strip()}"
|
||||
|
||||
|
||||
def _transcript(record: dict, index: int) -> str:
|
||||
m = record.get("metrics", {}) or {}
|
||||
correct = record.get("correct")
|
||||
kept = record.get("kept")
|
||||
flags = []
|
||||
if correct is not None:
|
||||
flags.append("correct" if correct else "WRONG")
|
||||
if kept is not None and kept:
|
||||
flags.append("kept")
|
||||
head = (
|
||||
f"### [{index}] task={record.get('task_id')} domain={record.get('domain')}"
|
||||
+ (f" {' '.join(flags)}" if flags else "")
|
||||
+ f"\n cost=${m.get('cost_usd')} tokens={m.get('tokens')} "
|
||||
f"tool_calls={m.get('num_tool_calls')} turns={m.get('num_turns')} "
|
||||
f"reward={record.get('reward')}"
|
||||
)
|
||||
anon_map = m.get("anon_map")
|
||||
body = [
|
||||
_render_turn(t, anon_map)
|
||||
for t in record.get("conversations", [])
|
||||
if str(t.get("role", "")).lower() != "system"
|
||||
]
|
||||
return head + "\n\n" + "\n\n".join(body)
|
||||
|
||||
|
||||
def _system_header(records: List[dict]) -> str:
|
||||
"""The system message is identical across records — render it once."""
|
||||
for r in records:
|
||||
for t in r.get("conversations", []):
|
||||
if str(t.get("role", "")).lower() == "system":
|
||||
return (
|
||||
"=" * 80 + "\nSHARED SYSTEM PROMPT + TOOL CATALOG "
|
||||
"(identical for every record below)\n" + "=" * 80
|
||||
+ f"\n{t.get('content', '').strip()}\n"
|
||||
)
|
||||
return ""
|
||||
|
||||
|
||||
_HTML_CSS = """
|
||||
body{font:14px/1.5 system-ui,sans-serif;margin:0;background:#f5f5f7;color:#1d1d1f}
|
||||
header{position:sticky;top:0;background:#1d1d1f;color:#fff;padding:12px 20px;z-index:9}
|
||||
header b{font-size:16px}
|
||||
.rec{background:#fff;margin:14px 20px;border-radius:10px;box-shadow:0 1px 3px #0002;overflow:hidden}
|
||||
.rec>summary{cursor:pointer;padding:10px 16px;font-weight:600;list-style:none;display:flex;gap:14px;align-items:center;flex-wrap:wrap}
|
||||
.rec>summary::-webkit-details-marker{display:none}
|
||||
.rec>summary:hover{background:#fafafa}
|
||||
.body{padding:6px 16px 16px}
|
||||
.turn{margin:10px 0;padding:10px 14px;border-radius:8px;white-space:pre-wrap;word-break:break-word}
|
||||
.user{background:#e8f0fe}
|
||||
.assistant{background:#f1f8e9}
|
||||
.call{background:#fff3e0}
|
||||
.call .cname{font-family:ui-monospace,monospace;font-weight:700;color:#b25b00;margin-bottom:6px}
|
||||
.call .cname .real{background:#b25b00;color:#fff;padding:1px 7px;border-radius:10px;font-size:12px;margin-left:4px}
|
||||
.call .arg{margin:6px 0 0 14px}
|
||||
.call .arg b{font-family:ui-monospace,monospace;color:#7a4500}
|
||||
.call .arg pre{white-space:pre-wrap;word-break:break-word;margin:3px 0 0;padding:8px 10px;background:#fff;border:1px solid #f0d9b8;border-radius:6px;font-size:13px}
|
||||
.tool details{background:#fafafa;border:1px solid #eee;border-radius:8px;padding:6px 10px;margin:10px 0}
|
||||
.tool summary{cursor:pointer;color:#555;font-weight:600}
|
||||
.tool pre{white-space:pre-wrap;word-break:break-word;margin:8px 0 0;font-size:13px}
|
||||
.role{font-size:11px;text-transform:uppercase;letter-spacing:.5px;color:#888;margin-bottom:4px}
|
||||
.pill{font-size:12px;padding:2px 8px;border-radius:10px;font-weight:600}
|
||||
.ok{background:#d7f5dd;color:#1a7f37}.wrong{background:#ffe0e0;color:#c0392b}
|
||||
.kept{background:#fff0c2;color:#9a6700}.dom{background:#eee;color:#444}
|
||||
.meta{font-weight:400;color:#888;font-size:12px}
|
||||
#sys details{margin:14px 20px;background:#fff;border-radius:10px;padding:10px 16px}
|
||||
#sys pre{white-space:pre-wrap;word-break:break-word;font-size:12px;color:#444}
|
||||
"""
|
||||
|
||||
|
||||
def _esc(s: str) -> str:
|
||||
return _htmllib.escape(s or "")
|
||||
|
||||
|
||||
def _html_turns(record: dict) -> str:
|
||||
anon_map = (record.get("metrics", {}) or {}).get("anon_map") or {}
|
||||
out = []
|
||||
for t in record.get("conversations", []):
|
||||
role = str(t.get("role", "")).lower()
|
||||
content = t.get("content", "") or ""
|
||||
if role == "system":
|
||||
continue
|
||||
if role == "user":
|
||||
out.append(f'<div class="turn user"><div class="role">user</div>{_esc(content.strip())}</div>')
|
||||
elif role in ("tool", "function"):
|
||||
name = _esc(t.get("name", "tool"))
|
||||
out.append(
|
||||
f'<div class="tool"><details><summary>tool result · {name} '
|
||||
f'({len(content)} chars)</summary><pre>{_esc(content.strip())}</pre></details></div>'
|
||||
)
|
||||
elif role == "assistant":
|
||||
prose = _TOOL_CALL_RE.sub("", content).strip()
|
||||
if prose:
|
||||
out.append(f'<div class="turn assistant"><div class="role">assistant</div>{_esc(prose)}</div>')
|
||||
for name, args in _parse_tool_calls(content):
|
||||
real = anon_map.get(name)
|
||||
label = (f'{_esc(str(name))} <span class="real">→ {_esc(str(real))}</span>'
|
||||
if real else _esc(str(name)))
|
||||
rows = [f'<div class="cname">→ {label}</div>']
|
||||
if isinstance(args, dict):
|
||||
for k, v in args.items():
|
||||
val = v if isinstance(v, str) else json.dumps(v, ensure_ascii=False, indent=2)
|
||||
rows.append(f'<div class="arg"><b>{_esc(str(k))}</b><pre>{_esc(val)}</pre></div>')
|
||||
out.append(f'<div class="turn call"><div class="role">tool call</div>{"".join(rows)}</div>')
|
||||
return "".join(out)
|
||||
|
||||
|
||||
def _html_record(record: dict, index: int) -> str:
|
||||
m = record.get("metrics", {}) or {}
|
||||
pills = [f'<span class="pill dom">{_esc(str(record.get("domain")))}</span>']
|
||||
correct = record.get("correct")
|
||||
if correct is not None:
|
||||
pills.append('<span class="pill ok">correct</span>' if correct
|
||||
else '<span class="pill wrong">wrong</span>')
|
||||
if record.get("kept"):
|
||||
pills.append('<span class="pill kept">kept</span>')
|
||||
meta = (f'<span class="meta">cost ${m.get("cost_usd")} · {m.get("tokens")} tok · '
|
||||
f'{m.get("num_tool_calls")} calls · {m.get("num_turns")} turns</span>')
|
||||
summary = (f'<span>#{index}</span><span>{_esc(str(record.get("task_id")))}</span>'
|
||||
+ "".join(pills) + meta)
|
||||
return (f'<details class="rec"><summary>{summary}</summary>'
|
||||
f'<div class="body">{_html_turns(record)}</div></details>')
|
||||
|
||||
|
||||
def _html_doc(records: List[dict], title: str) -> str:
|
||||
sys_txt = ""
|
||||
for r in records:
|
||||
for t in r.get("conversations", []):
|
||||
if str(t.get("role", "")).lower() == "system":
|
||||
sys_txt = t.get("content", "")
|
||||
break
|
||||
if sys_txt:
|
||||
break
|
||||
recs_html = "".join(_html_record(r, i + 1) for i, r in enumerate(records))
|
||||
return (
|
||||
f"<!doctype html><html><head><meta charset='utf-8'><title>{_esc(title)}</title>"
|
||||
f"<style>{_HTML_CSS}</style></head><body>"
|
||||
f"<header><b>{_esc(title)}</b> {len(records)} records "
|
||||
f"· click a row to expand</header>"
|
||||
f"<div id='sys'><details><summary style='cursor:pointer;font-weight:600;padding:4px'>"
|
||||
f"Shared system prompt + tool catalog (same for all)</summary>"
|
||||
f"<pre>{_esc(sys_txt.strip())}</pre></details></div>"
|
||||
f"{recs_html}</body></html>"
|
||||
)
|
||||
|
||||
|
||||
def render(jsonl_path: str) -> dict:
|
||||
src = Path(jsonl_path)
|
||||
records: List[dict] = []
|
||||
with src.open() as fh:
|
||||
for line in fh:
|
||||
line = line.strip()
|
||||
if line:
|
||||
records.append(json.loads(line))
|
||||
|
||||
pretty = src.with_suffix(".pretty.json")
|
||||
pretty.write_text(json.dumps(records, indent=2, ensure_ascii=False))
|
||||
|
||||
txt = src.with_suffix(".txt")
|
||||
banner = (
|
||||
f"ORCHESTRATOR SFT TRANSCRIPT — {src.name}\n{len(records)} records\n\n"
|
||||
)
|
||||
sep = f"\n\n{_RULE}\n\n"
|
||||
parts = [banner + _system_header(records)]
|
||||
parts.append(sep.join(_transcript(r, i + 1) for i, r in enumerate(records)))
|
||||
txt.write_text("\n\n".join(parts))
|
||||
|
||||
htmlp = src.with_suffix(".html")
|
||||
htmlp.write_text(_html_doc(records, src.name))
|
||||
|
||||
return {"records": len(records), "pretty": str(pretty),
|
||||
"txt": str(txt), "html": str(htmlp)}
|
||||
|
||||
|
||||
def main(argv: List[str]) -> int:
|
||||
if not argv:
|
||||
print(__doc__)
|
||||
return 2
|
||||
info = render(argv[0])
|
||||
print(json.dumps(info, indent=2))
|
||||
return 0
|
||||
|
||||
|
||||
if __name__ == "__main__":
|
||||
raise SystemExit(main(sys.argv[1:]))
|
||||
@@ -0,0 +1,338 @@
|
||||
#!/usr/bin/env python
|
||||
"""Full-parameter SFT for the orchestrator policy via FSDP (multi-GPU).
|
||||
|
||||
Launch with accelerate so the model + optimizer shard across GPUs (an 8-9B full
|
||||
fine-tune won't fit one L40S otherwise):
|
||||
|
||||
accelerate launch --config_file <fsdp.yaml> \
|
||||
scripts/orchestrator/run_sft_fsdp.py \
|
||||
--data data/runs/<...>/data.jsonl [--data <more> ...] \
|
||||
--variant correct --model Qwen/Qwen3.5-9B \
|
||||
--out checkpoints/sft_correct --epochs 3 --batch-size 8 --grad-accum 8 \
|
||||
--wandb-project orchestrator-sft
|
||||
|
||||
--variant selects which trajectories to train on:
|
||||
all -> every record
|
||||
correct -> only records the verifier marked correct
|
||||
correct_routed -> only correct records that delegated to a model expert
|
||||
|
||||
By default (--require-clean) records explicitly marked clean=False by the clean
|
||||
gate (bloated / garbled / unrouted) are also dropped; rows missing the flag
|
||||
(legacy data) are kept. Pass --no-require-clean to skip this filter.
|
||||
|
||||
Reuses the conversation->tokens + assistant-only masking from sft_tokenize.py.
|
||||
On these no-NVLink L40S, the accelerate/FSDP NCCL env (NCCL_P2P_DISABLE=1 etc.)
|
||||
must be set by the launcher.
|
||||
"""
|
||||
from __future__ import annotations
|
||||
|
||||
import argparse
|
||||
import json
|
||||
import math
|
||||
import os
|
||||
import re
|
||||
import sys
|
||||
import time
|
||||
from pathlib import Path
|
||||
|
||||
# reuse the data builder from the LoRA launcher
|
||||
sys.path.insert(0, os.path.dirname(os.path.abspath(__file__)))
|
||||
from sft_tokenize import build_example # noqa: E402
|
||||
|
||||
MODEL_EXPERTS = {"gpt_5_5", "claude_opus_4_8", "qwen3_5_9b", "qwen3_6_27b_fp8",
|
||||
"qwen3_5_122b_a10b_fp8", "qwen3_5_397b_a17b_fp8"}
|
||||
|
||||
|
||||
def record_is_correct(r: dict) -> bool:
|
||||
c = r.get("correct")
|
||||
if c is None:
|
||||
c = r.get("metrics", {}).get("correct")
|
||||
return bool(c)
|
||||
|
||||
|
||||
def record_routed_to_expert(r: dict) -> bool:
|
||||
"""True if the trajectory called a model expert (resolving anon labels)."""
|
||||
amap = r.get("metrics", {}).get("anon_map", {})
|
||||
names = re.findall(r'"name"\s*:\s*"([a-z0-9_]+)"', json.dumps(r.get("conversations", [])))
|
||||
for n in names:
|
||||
real = amap.get(n, n)
|
||||
if real in MODEL_EXPERTS:
|
||||
return True
|
||||
return False
|
||||
|
||||
|
||||
def record_clean_flag(r: dict):
|
||||
"""The record's `clean` flag: True/False, or None if absent (legacy data).
|
||||
|
||||
None means the clean gate never ran on this row, so we can't judge it -> we
|
||||
keep it (legacy behavior) rather than silently dropping older datasets.
|
||||
"""
|
||||
c = r.get("clean")
|
||||
if c is None:
|
||||
c = r.get("metrics", {}).get("clean")
|
||||
return c
|
||||
|
||||
|
||||
def select(records, variant: str, require_clean: bool = True):
|
||||
"""Filter by --variant, then (default) drop records explicitly marked unclean.
|
||||
|
||||
require_clean drops rows whose `clean` flag is False. Rows missing the flag
|
||||
(legacy) are kept regardless, matching pre-clean-gate behavior.
|
||||
"""
|
||||
if variant == "all":
|
||||
base = list(records)
|
||||
elif variant == "correct":
|
||||
base = [r for r in records if record_is_correct(r)]
|
||||
elif variant == "correct_routed":
|
||||
base = [r for r in records if record_is_correct(r) and record_routed_to_expert(r)]
|
||||
else:
|
||||
raise ValueError(f"unknown variant {variant}")
|
||||
if not require_clean:
|
||||
return base
|
||||
return [r for r in base if record_clean_flag(r) is not False]
|
||||
|
||||
|
||||
def main() -> int:
|
||||
p = argparse.ArgumentParser()
|
||||
p.add_argument("--data", action="append", required=True, help="JSONL(s) (repeatable)")
|
||||
p.add_argument("--val-data", default=None,
|
||||
help="Held-out JSONL for val-loss (excluded from --data). Eval'd per epoch.")
|
||||
p.add_argument("--variant", choices=["all", "correct", "correct_routed"], default="correct")
|
||||
p.add_argument("--require-clean", action=argparse.BooleanOptionalAction, default=True,
|
||||
help="Drop records whose `clean` flag is False (bloated / garbled / "
|
||||
"unrouted). Rows missing the flag (legacy data) are kept. "
|
||||
"Use --no-require-clean to disable.")
|
||||
p.add_argument("--model", default="Qwen/Qwen3.5-9B")
|
||||
p.add_argument("--out", required=True)
|
||||
p.add_argument("--epochs", type=float, default=3.0)
|
||||
p.add_argument("--batch-size", type=int, default=8)
|
||||
p.add_argument("--grad-accum", type=int, default=8)
|
||||
p.add_argument("--lr", type=float, default=1e-5)
|
||||
p.add_argument("--max-seq", type=int, default=8192)
|
||||
p.add_argument("--supervise-last-only", action="store_true",
|
||||
help="Legacy: supervise ONLY the final assistant turn. Default "
|
||||
"(off) supervises every assistant turn incl. routing.")
|
||||
p.add_argument("--warmup-ratio", type=float, default=0.03)
|
||||
p.add_argument("--seed", type=int, default=42)
|
||||
p.add_argument("--wandb-project", default="orchestrator-sft")
|
||||
p.add_argument("--wandb-name", default="")
|
||||
args = p.parse_args()
|
||||
|
||||
import random
|
||||
from datetime import timedelta
|
||||
import torch
|
||||
from torch.utils.data import DataLoader
|
||||
from accelerate import Accelerator
|
||||
from accelerate.utils import set_seed, InitProcessGroupKwargs
|
||||
from transformers import AutoModelForCausalLM, AutoTokenizer
|
||||
|
||||
set_seed(args.seed)
|
||||
# 30-min collective timeout: a slow first-step shard gather / graph build on a
|
||||
# contended node was tripping the default 10-min NCCL watchdog (job 3892 died
|
||||
# at step 1). This only delays a *real* hang's crash; it fixes spurious ones.
|
||||
pg_timeout_min = int(os.environ.get("SFT_PG_TIMEOUT_MIN", "30"))
|
||||
accelerator = Accelerator(
|
||||
gradient_accumulation_steps=args.grad_accum,
|
||||
kwargs_handlers=[InitProcessGroupKwargs(timeout=timedelta(minutes=pg_timeout_min))])
|
||||
is_main = accelerator.is_main_process
|
||||
|
||||
def log(m):
|
||||
if is_main:
|
||||
print(f"[fsdp-sft] {time.strftime('%H:%M:%S')} {m}", flush=True)
|
||||
|
||||
tok = AutoTokenizer.from_pretrained(args.model, trust_remote_code=True)
|
||||
if tok.pad_token_id is None:
|
||||
tok.pad_token = tok.eos_token
|
||||
|
||||
# ---- data ----
|
||||
records = []
|
||||
for f in args.data:
|
||||
records += [json.loads(l) for l in open(f) if l.strip()]
|
||||
sel = select(records, args.variant, require_clean=args.require_clean)
|
||||
log(f"variant={args.variant}: {len(sel)}/{len(records)} records")
|
||||
if args.require_clean:
|
||||
pre_clean = select(records, args.variant, require_clean=False)
|
||||
n_unclean = len(pre_clean) - len(sel)
|
||||
log(f"require_clean=True: dropped {n_unclean} unclean rows "
|
||||
f"({len(pre_clean)}->{len(sel)}); use --no-require-clean to keep them")
|
||||
else:
|
||||
log("require_clean=False: NOT filtering on `clean` flag")
|
||||
examples = []
|
||||
for r in sel:
|
||||
ex = build_example(tok, r.get("conversations", []), args.max_seq,
|
||||
supervise_all_turns=not args.supervise_last_only)
|
||||
if ex:
|
||||
examples.append(ex)
|
||||
log(f"built {len(examples)} training examples")
|
||||
if not examples:
|
||||
log("no examples; abort")
|
||||
return 1
|
||||
random.Random(args.seed).shuffle(examples)
|
||||
|
||||
def collate(batch):
|
||||
maxlen = max(len(b["input_ids"]) for b in batch)
|
||||
ii, lab, am = [], [], []
|
||||
for b in batch:
|
||||
pad = maxlen - len(b["input_ids"])
|
||||
ii.append(b["input_ids"] + [tok.pad_token_id] * pad)
|
||||
lab.append(b["labels"] + [-100] * pad)
|
||||
am.append([1] * len(b["input_ids"]) + [0] * pad)
|
||||
return (torch.tensor(ii), torch.tensor(lab), torch.tensor(am))
|
||||
|
||||
dl = DataLoader(examples, batch_size=args.batch_size, shuffle=True, collate_fn=collate)
|
||||
|
||||
# ---- held-out val set (leak-free; excluded from --data upstream) ----
|
||||
val_dl = None
|
||||
if args.val_data:
|
||||
vrecs = [json.loads(l) for l in open(args.val_data) if l.strip()]
|
||||
vex = []
|
||||
for r in vrecs:
|
||||
ex = build_example(tok, r.get("conversations", []), args.max_seq,
|
||||
supervise_all_turns=not args.supervise_last_only)
|
||||
if ex:
|
||||
vex.append(ex)
|
||||
log(f"val: {len(vex)} examples from {args.val_data}")
|
||||
if vex:
|
||||
val_dl = DataLoader(vex, batch_size=args.batch_size, shuffle=False, collate_fn=collate)
|
||||
|
||||
model = AutoModelForCausalLM.from_pretrained(
|
||||
args.model, dtype=torch.bfloat16, trust_remote_code=True,
|
||||
attn_implementation="sdpa")
|
||||
if os.environ.get("SFT_NO_GRAD_CKPT") == "1":
|
||||
log("gradient checkpointing DISABLED (SFT_NO_GRAD_CKPT=1)")
|
||||
else:
|
||||
model.gradient_checkpointing_enable(gradient_checkpointing_kwargs={"use_reentrant": False})
|
||||
model.config.use_cache = False
|
||||
opt = torch.optim.AdamW(model.parameters(), lr=args.lr, weight_decay=0.0)
|
||||
|
||||
model, opt, dl = accelerator.prepare(model, opt, dl)
|
||||
if val_dl is not None:
|
||||
val_dl = accelerator.prepare(val_dl)
|
||||
|
||||
@torch.no_grad()
|
||||
def evaluate():
|
||||
# Corpus-level mean token loss over the holdout. FSDP needs every rank to
|
||||
# run the forward collectively, so each rank evals its shard and we reduce
|
||||
# token-weighted sums (not a mean-of-batch-means, which would mis-weight).
|
||||
model.eval()
|
||||
tot_loss = torch.zeros((), device=accelerator.device)
|
||||
tot_tok = torch.zeros((), device=accelerator.device)
|
||||
for ii, lab, am in val_dl:
|
||||
out = model(input_ids=ii, attention_mask=am, labels=lab)
|
||||
ntok = (lab != -100).sum()
|
||||
tot_loss += out.loss.detach() * ntok
|
||||
tot_tok += ntok
|
||||
tot_loss = accelerator.reduce(tot_loss, reduction="sum")
|
||||
tot_tok = accelerator.reduce(tot_tok, reduction="sum")
|
||||
model.train()
|
||||
return (tot_loss / tot_tok.clamp(min=1)).item()
|
||||
|
||||
steps_per_epoch = math.ceil(len(dl) / args.grad_accum)
|
||||
total_steps = max(1, int(steps_per_epoch * args.epochs))
|
||||
warmup = max(1, int(total_steps * args.warmup_ratio))
|
||||
|
||||
def lr_at(s):
|
||||
if s < warmup:
|
||||
return s / warmup
|
||||
return 0.5 * (1 + math.cos(math.pi * (s - warmup) / max(1, total_steps - warmup)))
|
||||
|
||||
use_wandb = False
|
||||
if is_main:
|
||||
try:
|
||||
import wandb
|
||||
base_name = args.wandb_name or f"{Path(args.model).name}-{args.variant}-fsdp"
|
||||
run_name = f"{base_name}-{time.strftime('%m%d-%H%M', time.gmtime())}"
|
||||
wandb.init(project=args.wandb_project,
|
||||
name=run_name,
|
||||
config=vars(args) | {"n_examples": len(examples), "total_steps": total_steps})
|
||||
use_wandb = True
|
||||
except Exception as e:
|
||||
log(f"wandb off ({e})")
|
||||
|
||||
log(f"total_steps={total_steps} steps/epoch={steps_per_epoch} warmup={warmup}")
|
||||
|
||||
def _save_ckpt(out_path):
|
||||
# Full-model checkpoint (gathers FSDP shards to rank0). Called per-epoch
|
||||
# so a long run leaves usable partial models + lets us eval intermediate
|
||||
# checkpoints (e.g. epoch1/epoch2) for the data-scaling sweep.
|
||||
accelerator.wait_for_everyone()
|
||||
unwrapped = accelerator.unwrap_model(model)
|
||||
p = Path(out_path)
|
||||
if is_main:
|
||||
p.mkdir(parents=True, exist_ok=True)
|
||||
tok.save_pretrained(str(p))
|
||||
unwrapped.save_pretrained(
|
||||
str(p), is_main_process=is_main,
|
||||
save_function=accelerator.save,
|
||||
state_dict=accelerator.get_state_dict(model))
|
||||
if is_main:
|
||||
log(f"checkpoint saved -> {p}")
|
||||
|
||||
model.train()
|
||||
gstep = 0
|
||||
t0 = time.time()
|
||||
if val_dl is not None:
|
||||
vloss = evaluate()
|
||||
log(f"val/loss (baseline, step 0) = {vloss:.4f}")
|
||||
if use_wandb:
|
||||
wandb.log({"val/loss": vloss}, step=0)
|
||||
for epoch in range(math.ceil(args.epochs)):
|
||||
for ii, lab, am in dl:
|
||||
with accelerator.accumulate(model):
|
||||
out = model(input_ids=ii, attention_mask=am, labels=lab)
|
||||
accelerator.backward(out.loss)
|
||||
if accelerator.sync_gradients:
|
||||
lr = args.lr * lr_at(gstep)
|
||||
for g in opt.param_groups:
|
||||
g["lr"] = lr
|
||||
accelerator.clip_grad_norm_(model.parameters(), 1.0)
|
||||
opt.step()
|
||||
opt.zero_grad()
|
||||
if accelerator.sync_gradients:
|
||||
gstep += 1
|
||||
# accelerator.gather() is a COLLECTIVE — EVERY rank must call it.
|
||||
# Guarding it behind `is_main` made only rank0 run the [1]-float
|
||||
# all-gather while ranks 1-3 marched on to the next step's FSDP
|
||||
# param all-gather, desyncing the process group and hanging the job
|
||||
# at the step1->step2 boundary (NCCL collective-shape mismatch).
|
||||
loss = accelerator.gather(out.loss.detach()).mean().item()
|
||||
if is_main:
|
||||
log(f"step {gstep}/{total_steps} loss={loss:.4f} lr={lr:.2e}")
|
||||
if use_wandb:
|
||||
wandb.log({"train/loss": loss, "train/lr": lr}, step=gstep)
|
||||
if gstep >= total_steps:
|
||||
break
|
||||
_save_ckpt(Path(args.out) / f"epoch{epoch + 1}")
|
||||
if val_dl is not None:
|
||||
vloss = evaluate()
|
||||
log(f"val/loss (epoch {epoch + 1}) = {vloss:.4f}")
|
||||
if use_wandb:
|
||||
wandb.log({"val/loss": vloss, "epoch": epoch + 1}, step=gstep)
|
||||
if gstep >= total_steps:
|
||||
break
|
||||
|
||||
accelerator.wait_for_everyone()
|
||||
log("saving full model...")
|
||||
unwrapped = accelerator.unwrap_model(model)
|
||||
out_dir = Path(args.out)
|
||||
if is_main:
|
||||
out_dir.mkdir(parents=True, exist_ok=True)
|
||||
tok.save_pretrained(str(out_dir))
|
||||
unwrapped.save_pretrained(
|
||||
str(out_dir), is_main_process=is_main,
|
||||
save_function=accelerator.save,
|
||||
state_dict=accelerator.get_state_dict(model))
|
||||
if is_main:
|
||||
log(f"saved -> {out_dir}")
|
||||
if use_wandb:
|
||||
try:
|
||||
wandb.finish()
|
||||
except Exception:
|
||||
pass
|
||||
print("FSDP_SFT_DONE " + json.dumps({"variant": args.variant, "steps": gstep,
|
||||
"wall_s": round(time.time() - t0, 1), "out": str(out_dir)}))
|
||||
return 0
|
||||
|
||||
|
||||
if __name__ == "__main__":
|
||||
raise SystemExit(main())
|
||||
@@ -0,0 +1,13 @@
|
||||
compute_environment: LOCAL_MACHINE
|
||||
distributed_type: FSDP
|
||||
mixed_precision: bf16
|
||||
num_machines: 1
|
||||
num_processes: 4
|
||||
fsdp_config:
|
||||
fsdp_sharding_strategy: FULL_SHARD
|
||||
fsdp_auto_wrap_policy: SIZE_BASED_WRAP
|
||||
fsdp_min_num_params: 100000000
|
||||
fsdp_state_dict_type: FULL_STATE_DICT
|
||||
fsdp_use_orig_params: true
|
||||
fsdp_cpu_ram_efficient_loading: false
|
||||
fsdp_sync_module_states: false
|
||||
@@ -0,0 +1,148 @@
|
||||
#!/usr/bin/env python
|
||||
"""Conversation->tokens + assistant-only masking shared by the SFT trainers.
|
||||
|
||||
Extracted from the old LoRA trainer so ``run_sft_fsdp.py`` (full-parameter FSDP,
|
||||
the path we actually use) doesn't depend on it. ``build_example`` tokenizes one
|
||||
``conversations`` record and returns ``input_ids`` + ``labels`` with everything
|
||||
but the supervised assistant turns masked to -100.
|
||||
"""
|
||||
from __future__ import annotations
|
||||
|
||||
from typing import Any, Dict, List, Optional
|
||||
|
||||
|
||||
def normalize_messages(conversations: List[Dict[str, Any]]) -> List[Dict[str, str]]:
|
||||
"""Map raw conversation turns to {role, content} with role in
|
||||
system/user/assistant. tool turns fold into user (matches the native
|
||||
OrchestratorSFTDataset fallback)."""
|
||||
msgs: List[Dict[str, str]] = []
|
||||
for turn in conversations:
|
||||
role = turn.get("role") or turn.get("from", "")
|
||||
content = turn.get("content") or turn.get("value", "") or ""
|
||||
if role in ("human", "user"):
|
||||
role = "user"
|
||||
elif role in ("gpt", "assistant"):
|
||||
role = "assistant"
|
||||
elif role == "tool":
|
||||
content = f"[Tool '{turn.get('name', 'tool')}' returned]: {content}"
|
||||
role = "user"
|
||||
if role in ("system", "user", "assistant"):
|
||||
msgs.append({"role": role, "content": str(content)})
|
||||
return msgs
|
||||
|
||||
|
||||
def _lcp_len(a: List[int], b: List[int]) -> int:
|
||||
"""Length of the longest common prefix of two token-id lists."""
|
||||
n = 0
|
||||
for x, y in zip(a, b):
|
||||
if x != y:
|
||||
break
|
||||
n += 1
|
||||
return n
|
||||
|
||||
|
||||
def _chatml_assistant_spans(ids: List[int], tok) -> Optional[List[tuple]]:
|
||||
"""``(start, end)`` token spans of each assistant turn's content in a
|
||||
ChatML-rendered sequence, located by ``<|im_start|>assistant ... <|im_end|>``
|
||||
markers (the ``<|im_end|>`` is included so the model learns to stop).
|
||||
|
||||
Operates on the FINAL token stream, so it is correct even when the chat
|
||||
template strips prior-turn ``<think>`` reasoning from history (Qwen3.x does),
|
||||
which breaks any incremental re-rendering / prefix approach. Returns None for
|
||||
non-ChatML templates (e.g. gemma) so the caller can fall back."""
|
||||
im_start = tok.convert_tokens_to_ids("<|im_start|>")
|
||||
im_end = tok.convert_tokens_to_ids("<|im_end|>")
|
||||
asst = tok.convert_tokens_to_ids("assistant")
|
||||
unk = tok.unk_token_id
|
||||
if im_start is None or im_end is None or im_start == unk or im_end == unk:
|
||||
return None
|
||||
nl = tok.convert_tokens_to_ids("Ċ") # the '\n' after the role header
|
||||
spans: List[tuple] = []
|
||||
i, n = 0, len(ids)
|
||||
while i < n:
|
||||
if ids[i] == im_start and i + 1 < n and ids[i + 1] == asst:
|
||||
j = i + 2
|
||||
if j < n and ids[j] == nl:
|
||||
j += 1
|
||||
k = j
|
||||
while k < n and ids[k] != im_end:
|
||||
k += 1
|
||||
end = min(k + 1, n)
|
||||
if end > j:
|
||||
spans.append((j, end))
|
||||
i = end
|
||||
else:
|
||||
i += 1
|
||||
return spans
|
||||
|
||||
|
||||
def build_example(
|
||||
tok,
|
||||
conversations,
|
||||
max_seq: int,
|
||||
supervise_all_turns: bool = True,
|
||||
) -> Optional[Dict[str, List[int]]]:
|
||||
"""Tokenize one record -> input_ids + labels.
|
||||
|
||||
``supervise_all_turns=True`` (default): supervise EVERY assistant turn — the
|
||||
intermediate routing ``<tool_call>`` turns AND the final answer; only
|
||||
system/user/tool tokens are masked. This teaches the model to actually route,
|
||||
not just synthesize the last answer given someone else's tool calls.
|
||||
|
||||
``supervise_all_turns=False`` (legacy): supervise only the final assistant
|
||||
turn; everything before it is masked to -100.
|
||||
|
||||
Assistant spans are found by the ChatML turn markers on the rendered stream
|
||||
(robust to Qwen3's prior-turn think-stripping). For non-ChatML templates
|
||||
(gemma) we fall back to the proven last-turn longest-common-prefix boundary;
|
||||
all-turns supervision there needs per-template markers and is not yet wired,
|
||||
so it degrades to last-turn. Records whose last turn isn't assistant, that
|
||||
supervise nothing, or whose final turn alone overflows max_seq are skipped."""
|
||||
msgs = normalize_messages(conversations)
|
||||
if len(msgs) < 2 or msgs[-1]["role"] != "assistant":
|
||||
return None
|
||||
try:
|
||||
full_text = tok.apply_chat_template(msgs, tokenize=False,
|
||||
add_generation_prompt=False)
|
||||
full = tok(full_text, add_special_tokens=False)["input_ids"]
|
||||
except Exception:
|
||||
return None
|
||||
if not full:
|
||||
return None
|
||||
|
||||
spans = _chatml_assistant_spans(full, tok)
|
||||
if spans:
|
||||
selected = spans if supervise_all_turns else spans[-1:]
|
||||
labels = [-100] * len(full)
|
||||
for (s, e) in selected:
|
||||
for k in range(s, e):
|
||||
labels[k] = full[k]
|
||||
last_start = selected[-1][0]
|
||||
else:
|
||||
# Non-ChatML template: proven last-turn boundary via longest-common-prefix
|
||||
# (robust to templates whose add_generation_prompt emits extra preamble).
|
||||
try:
|
||||
prompt_text = tok.apply_chat_template(msgs[:-1], tokenize=False,
|
||||
add_generation_prompt=True)
|
||||
prompt = tok(prompt_text, add_special_tokens=False)["input_ids"]
|
||||
except Exception:
|
||||
return None
|
||||
boundary = _lcp_len(prompt, full)
|
||||
if boundary >= len(full):
|
||||
return None
|
||||
labels = list(full)
|
||||
for i in range(min(boundary, len(labels))):
|
||||
labels[i] = -100
|
||||
last_start = boundary
|
||||
|
||||
if all(v == -100 for v in labels): # nothing left to supervise
|
||||
return None
|
||||
|
||||
# Keep the final assistant turn if too long: left-truncate from the front.
|
||||
if len(full) > max_seq:
|
||||
if len(full) - last_start >= max_seq: # final answer alone overflows -> drop
|
||||
return None
|
||||
cut = len(full) - max_seq
|
||||
full = full[cut:]
|
||||
labels = labels[cut:]
|
||||
return {"input_ids": full, "labels": labels}
|
||||
@@ -0,0 +1,361 @@
|
||||
#!/usr/bin/env python
|
||||
"""Upload orchestrator-SFT JSONL datasets to Braintrust so they're browsable/
|
||||
filterable in the UI (by area / difficulty / dataset / routed-model / correct).
|
||||
|
||||
Datasets land in the SAME Braintrust project as the rollout TRACES (the
|
||||
``research`` project) so a run's data + traces sit side by side. The target is
|
||||
resolved from ``OJ_BRAINTRUST_PROJECT_ID`` (default: the research project id);
|
||||
``--project`` overrides by NAME if you want a different project.
|
||||
|
||||
Each record -> a Braintrust dataset row:
|
||||
input = the problem/question
|
||||
expected = the FINAL_ANSWER value
|
||||
tags = [gen_model, domain, correct|incorrect, clean|dirty, kept|dropped]
|
||||
metadata = domain, area, difficulty, dataset, subsector, task_id, correct,
|
||||
clean, kept, reward, gen_model, orchestrator_model, num_tool_calls,
|
||||
n_turns, routed_models (real names)
|
||||
(the full conversation is kept under metadata.conversation for inspection)
|
||||
|
||||
The dataset itself carries run-level metadata (run_label, specific gen_model,
|
||||
git sha, config knobs, counts, domain distribution) so future experiments are
|
||||
distinguishable in the UI.
|
||||
|
||||
Usage (CLI):
|
||||
.venv/bin/python scripts/orchestrator/upload_to_braintrust.py \
|
||||
data/sft_variants/qwen_orch_train_0707.jsonl=qwen_train_0707 \
|
||||
data/sft_variants/qwen_orch_holdout_0707.jsonl=qwen_holdout_0707
|
||||
|
||||
# name is optional; defaults to {model_short}_{split}_{date}
|
||||
.venv/bin/python scripts/orchestrator/upload_to_braintrust.py \
|
||||
data/sft_variants/qwen_orch_holdout_0707.jsonl
|
||||
|
||||
Programmatic (used by the pipeline auto-upload hook in make_splits.py):
|
||||
from upload_to_braintrust import autoupload
|
||||
autoupload(["data/sft_variants/qwen_orch_train_0707.jsonl=qwen_train_0707"],
|
||||
run_label="...")
|
||||
"""
|
||||
import argparse
|
||||
import json
|
||||
import logging
|
||||
import os
|
||||
import re
|
||||
import subprocess
|
||||
from collections import Counter
|
||||
from datetime import datetime
|
||||
from pathlib import Path
|
||||
|
||||
logger = logging.getLogger(__name__)
|
||||
|
||||
# The `research` project — datasets go here, same place the rollout traces land.
|
||||
DEFAULT_PROJECT_ID = "c707124e-ce9d-4187-ad11-f49f19f777ad"
|
||||
|
||||
_FA = re.compile(r"(?im)FINAL[_\s]?ANSWER\s*:?")
|
||||
|
||||
|
||||
def _ordered_turns(convs, anon_map, orchestrator):
|
||||
"""Rebuild each turn as role/model/content, in that display order.
|
||||
|
||||
Braintrust sorts object keys alphabetically on write, which would push the
|
||||
(very long) `content` to the top. We prefix the keys (`1_role`, `2_model`,
|
||||
`3_content`) so they sort into role -> model -> content and the content
|
||||
renders last. `model` is the model behind the turn:
|
||||
* assistant turns -> the orchestrator model
|
||||
* tool turns -> the expert model that answered (de-anonymized)
|
||||
* system/user -> None
|
||||
"""
|
||||
out = []
|
||||
for c in convs:
|
||||
role = c.get("role")
|
||||
if role == "tool":
|
||||
model = anon_map.get(c.get("name")) or c.get("name")
|
||||
elif role == "assistant":
|
||||
model = orchestrator
|
||||
else:
|
||||
model = None
|
||||
out.append({"1_role": role, "2_model": model, "3_content": c.get("content")})
|
||||
return out
|
||||
|
||||
|
||||
def _dataset_desc(dataset):
|
||||
"""Fallback dataset blurb for records generated before the field existed."""
|
||||
try:
|
||||
from openjarvis.learning.intelligence.orchestrator.sft_data.reject_sample import (
|
||||
dataset_description,
|
||||
)
|
||||
return dataset_description(dataset)
|
||||
except Exception:
|
||||
return ""
|
||||
|
||||
|
||||
def _git_sha():
|
||||
try:
|
||||
return subprocess.check_output(
|
||||
["git", "rev-parse", "--short", "HEAD"],
|
||||
cwd=Path(__file__).resolve().parent, stderr=subprocess.DEVNULL,
|
||||
).decode().strip() or None
|
||||
except Exception:
|
||||
return None
|
||||
|
||||
|
||||
def _model_short(rows, path):
|
||||
"""Short family tag for the default dataset name (qwen / gemma / ...)."""
|
||||
stem = Path(path).stem.lower()
|
||||
for fam in ("qwen", "gemma"):
|
||||
if stem.startswith(fam) or fam in stem:
|
||||
return fam
|
||||
gm = next((r.get("gen_model") for r in rows if r.get("gen_model")), "") or ""
|
||||
gm = gm.lower()
|
||||
for fam in ("qwen", "gemma"):
|
||||
if fam in gm:
|
||||
return fam
|
||||
return "orch"
|
||||
|
||||
|
||||
def _split_of(path):
|
||||
stem = Path(path).stem.lower()
|
||||
for s in ("holdout", "train", "clean", "overfit", "partial", "8k", "4k", "2k", "1k"):
|
||||
if s in stem:
|
||||
return s
|
||||
return "data"
|
||||
|
||||
|
||||
def _default_name(rows, path):
|
||||
return f"{_model_short(rows, path)}_{_split_of(path)}_{datetime.now():%m%d}"
|
||||
|
||||
|
||||
def _config_from_env():
|
||||
"""Config knobs from the run env (set by build_orchestrator_sft.py). Omit unset."""
|
||||
cfg = {}
|
||||
for env_key, meta_key, cast in [
|
||||
("OJ_CFG_TEMPERATURE", "temperature", float),
|
||||
("OJ_CFG_MAX_TURNS", "max_turns", int),
|
||||
("OJ_CFG_ANONYMIZE", "anonymize", lambda v: v.strip().lower() in ("1", "true", "yes")),
|
||||
("OJ_CFG_REJECTION_ONLY", "rejection_only", lambda v: v.strip().lower() in ("1", "true", "yes")),
|
||||
]:
|
||||
v = os.getenv(env_key)
|
||||
if v not in (None, ""):
|
||||
try:
|
||||
cfg[meta_key] = cast(v)
|
||||
except Exception:
|
||||
cfg[meta_key] = v
|
||||
return cfg
|
||||
|
||||
|
||||
def _row(rec):
|
||||
convs = rec.get("conversations", [])
|
||||
user = next((c["content"] for c in convs if c["role"] == "user"), "")
|
||||
question = re.sub(r"^\s*Problem:\s*", "", user, count=1).strip()
|
||||
asst = [c["content"] for c in convs if c["role"] == "assistant"]
|
||||
fa = asst[-1] if asst else ""
|
||||
m = list(_FA.finditer(fa))
|
||||
model_answer = fa[m[-1].end():].strip() if m else fa.strip()
|
||||
# Gold reference the verifier graded against (stamped at generation time as
|
||||
# `gold_answer`). `expected` in Braintrust is the GOLD so the UI shows
|
||||
# gold-vs-model; the model's own answer goes to metadata.model_answer.
|
||||
gold = rec.get("gold_answer") or None
|
||||
has_gold = gold is not None
|
||||
am = (rec.get("metrics", {}) or {}).get("anon_map", {}) or {}
|
||||
routed = []
|
||||
for c in convs:
|
||||
if c["role"] == "assistant":
|
||||
for t in re.findall(r"<tool_call>(\{.*?\})</tool_call>", c["content"], re.DOTALL):
|
||||
try:
|
||||
real = am.get(json.loads(t).get("name"))
|
||||
if real:
|
||||
routed.append(real)
|
||||
except Exception:
|
||||
pass
|
||||
metrics = rec.get("metrics", {}) or {}
|
||||
gen_model = rec.get("gen_model")
|
||||
domain = rec.get("domain")
|
||||
correct = rec.get("correct")
|
||||
clean = rec.get("clean")
|
||||
kept = rec.get("kept")
|
||||
tags = [t for t in [
|
||||
gen_model,
|
||||
f"domain:{domain}" if domain else None,
|
||||
"correct" if correct else "incorrect",
|
||||
"clean" if clean else "dirty",
|
||||
"kept" if kept else "dropped",
|
||||
"gold" if has_gold else "no_gold",
|
||||
] if t]
|
||||
return {
|
||||
"input": question,
|
||||
"expected": gold,
|
||||
"tags": tags,
|
||||
"metadata": {
|
||||
"gold_answer": gold,
|
||||
"model_answer": model_answer,
|
||||
"has_gold": has_gold,
|
||||
"domain": domain,
|
||||
"area": rec.get("area"),
|
||||
"difficulty": rec.get("difficulty") or None,
|
||||
"dataset": rec.get("dataset"),
|
||||
"dataset_description": rec.get("dataset_description") or _dataset_desc(rec.get("dataset")),
|
||||
"subsector": rec.get("subsector"),
|
||||
"task_id": rec.get("task_id"),
|
||||
"correct": correct,
|
||||
"clean": clean,
|
||||
"kept": kept,
|
||||
"reward": rec.get("reward"),
|
||||
"gen_model": gen_model,
|
||||
"orchestrator_model": rec.get("orchestrator_model"),
|
||||
"num_tool_calls": metrics.get("num_tool_calls"),
|
||||
"n_turns": metrics.get("num_turns"),
|
||||
"routed_models": routed,
|
||||
"conversation": _ordered_turns(convs, am, rec.get("orchestrator_model")),
|
||||
},
|
||||
}
|
||||
|
||||
|
||||
def _dist(rows, key):
|
||||
"""{value: count} for a record field, most-common first, None/'' dropped."""
|
||||
c = Counter(r.get(key) for r in rows if (r.get(key) or "") != "")
|
||||
return {str(k): v for k, v in sorted(c.items(), key=lambda x: -x[1])}
|
||||
|
||||
|
||||
def _routed_dist(rows):
|
||||
"""Distribution of real (de-anonymized) expert models routed to."""
|
||||
c = Counter()
|
||||
for r in rows:
|
||||
am = (r.get("metrics", {}) or {}).get("anon_map", {}) or {}
|
||||
for cc in r.get("conversations", []):
|
||||
if cc.get("role") != "assistant":
|
||||
continue
|
||||
for t in re.findall(r"<tool_call>(\{.*?\})</tool_call>", cc.get("content", ""), re.DOTALL):
|
||||
try:
|
||||
real = am.get(json.loads(t).get("name"))
|
||||
if real:
|
||||
c[real] += 1
|
||||
except Exception:
|
||||
pass
|
||||
return {k: v for k, v in sorted(c.items(), key=lambda x: -x[1])}
|
||||
|
||||
|
||||
def _avg(rows, *path):
|
||||
vals = []
|
||||
for r in rows:
|
||||
v = r.get("metrics", {}) or {}
|
||||
for p in path:
|
||||
v = (v or {}).get(p) if isinstance(v, dict) else None
|
||||
if isinstance(v, (int, float)):
|
||||
vals.append(v)
|
||||
return round(sum(vals) / len(vals), 2) if vals else None
|
||||
|
||||
|
||||
def _dataset_metadata(rows, path, run_label, gen_model):
|
||||
n = len(rows)
|
||||
meta = {
|
||||
"model": _model_short(rows, path),
|
||||
"split": _split_of(path),
|
||||
"task": "orchestrator-SFT unified-tool routing traces",
|
||||
"date": "2026-07-07",
|
||||
"run_label": run_label,
|
||||
"gen_model": gen_model,
|
||||
"git_sha": _git_sha(),
|
||||
"source_file": str(path),
|
||||
"source_datasets": ["GeneralThought-430K-filtered", "OpenThoughts3-1.2M"],
|
||||
"task_mix": "balanced (GeneralThought + OpenThoughts code/math/science)",
|
||||
"filter": "correct + clean (rejects file-write echoes, garble, "
|
||||
"reasoning-degeneration, truncated tails, essays/markdown dumps)",
|
||||
"leak_free": "train/holdout disjoint by task_id",
|
||||
"n_total": n,
|
||||
"n_correct": sum(1 for r in rows if r.get("correct")),
|
||||
"n_clean": sum(1 for r in rows if r.get("clean")),
|
||||
"n_kept": sum(1 for r in rows if r.get("kept")),
|
||||
"n_with_gold": sum(1 for r in rows if (r.get("gold_answer") or "").strip()),
|
||||
"avg_tool_calls": _avg(rows, "num_tool_calls"),
|
||||
"avg_turns": _avg(rows, "num_turns"),
|
||||
"area_distribution": _dist(rows, "area"),
|
||||
"difficulty_distribution": _dist(rows, "difficulty"),
|
||||
"dataset_distribution": _dist(rows, "dataset"),
|
||||
"routed_model_distribution": _routed_dist(rows),
|
||||
}
|
||||
cfg = _config_from_env()
|
||||
if cfg:
|
||||
meta["config"] = cfg
|
||||
return {k: v for k, v in meta.items() if v is not None}
|
||||
|
||||
|
||||
def upload_dataset(path, name=None, *, project_id=None, project=None,
|
||||
run_label=None, gen_model=None, description=""):
|
||||
"""Upload one JSONL file as a Braintrust dataset. Returns (name, n_rows, url).
|
||||
|
||||
Target project: ``project`` (name) if given, else ``project_id`` /
|
||||
``OJ_BRAINTRUST_PROJECT_ID`` / the research default.
|
||||
"""
|
||||
import braintrust
|
||||
|
||||
rows = [json.loads(l) for l in open(path) if l.strip()]
|
||||
name = name or _default_name(rows, path)
|
||||
gen_model = gen_model or os.getenv("OJ_GEN_MODEL") or \
|
||||
next((r.get("gen_model") for r in rows if r.get("gen_model")), None)
|
||||
run_label = run_label or os.getenv("OJ_RUN_LABEL") or name
|
||||
|
||||
init_kwargs = {"name": name, "description": description or None,
|
||||
"metadata": _dataset_metadata(rows, path, run_label, gen_model)}
|
||||
if project:
|
||||
init_kwargs["project"] = project
|
||||
else:
|
||||
init_kwargs["project_id"] = project_id or os.getenv(
|
||||
"OJ_BRAINTRUST_PROJECT_ID", DEFAULT_PROJECT_ID)
|
||||
|
||||
ds = braintrust.init_dataset(**init_kwargs)
|
||||
for r in rows:
|
||||
ds.insert(**_row(r))
|
||||
ds.flush()
|
||||
summ = ds.summarize()
|
||||
url = getattr(summ, "dataset_url", None) or (project or init_kwargs.get("project_id"))
|
||||
return name, len(rows), url
|
||||
|
||||
|
||||
def autoupload(specs, *, run_label=None, gen_model=None, description=""):
|
||||
"""No-op-safe wrapper for pipeline hooks. Honors OJ_BRAINTRUST_AUTOUPLOAD
|
||||
(default ON) and never raises: any failure (missing key/pkg, network) is
|
||||
logged and swallowed so the data pipeline is never broken by telemetry."""
|
||||
if os.getenv("OJ_BRAINTRUST_AUTOUPLOAD", "1").strip().lower() in ("0", "false", "no", "off", ""):
|
||||
logger.info("[braintrust] autoupload disabled (OJ_BRAINTRUST_AUTOUPLOAD)")
|
||||
return
|
||||
if not os.getenv("BRAINTRUST_API_KEY"):
|
||||
logger.info("[braintrust] autoupload skipped — BRAINTRUST_API_KEY unset")
|
||||
return
|
||||
try:
|
||||
import braintrust # noqa: F401
|
||||
except Exception:
|
||||
logger.info("[braintrust] autoupload skipped — braintrust not installed")
|
||||
return
|
||||
for spec in specs:
|
||||
path, _, name = spec.partition("=")
|
||||
try:
|
||||
nm, n, url = upload_dataset(path, name or None, run_label=run_label,
|
||||
gen_model=gen_model, description=description)
|
||||
logger.info("[braintrust] uploaded %s: %d rows -> %s", nm, n, url)
|
||||
print(f"[braintrust] uploaded {nm}: {n} rows -> {url}")
|
||||
except Exception as exc:
|
||||
logger.warning("[braintrust] autoupload FAILED for %s (%s) — continuing", path, exc)
|
||||
print(f"[braintrust] autoupload FAILED for {path} ({exc}) — pipeline unaffected")
|
||||
|
||||
|
||||
def main():
|
||||
logging.basicConfig(level=logging.INFO, format="%(message)s")
|
||||
p = argparse.ArgumentParser()
|
||||
p.add_argument("--project", default=None,
|
||||
help="Target project by NAME (overrides OJ_BRAINTRUST_PROJECT_ID).")
|
||||
p.add_argument("--project-id", default=None,
|
||||
help="Target project by id (default: OJ_BRAINTRUST_PROJECT_ID or research).")
|
||||
p.add_argument("--run-label", default=None)
|
||||
p.add_argument("--gen-model", default=None, help="Specific gen model id override.")
|
||||
p.add_argument("--description", default="")
|
||||
p.add_argument("specs", nargs="+", help="path.jsonl[=dataset_name]")
|
||||
args = p.parse_args()
|
||||
for spec in args.specs:
|
||||
path, _, name = spec.partition("=")
|
||||
nm, n, url = upload_dataset(
|
||||
path, name or None, project_id=args.project_id, project=args.project,
|
||||
run_label=args.run_label, gen_model=args.gen_model,
|
||||
description=args.description)
|
||||
print(f"[braintrust] {nm}: {n} rows -> {url}")
|
||||
|
||||
|
||||
if __name__ == "__main__":
|
||||
raise SystemExit(main())
|
||||
@@ -0,0 +1,90 @@
|
||||
#!/bin/bash
|
||||
# Serve a trained gemma-4-26B-A4B SFT checkpoint (orchestrator) on 2xL40S (mkt1),
|
||||
# then run eval_orchestrator over benchmarks. Self-contained: serve -> health -> eval -> teardown.
|
||||
# Usage: sbatch eval_orch_gemma_matx.sbatch <1k|2k> <benchmarks> <N> <tag>
|
||||
#SBATCH --job-name=eval_gemma
|
||||
#SBATCH --account=mkt
|
||||
#SBATCH --partition=mkt
|
||||
#SBATCH -w mkt1
|
||||
#SBATCH --gres=gpu:2
|
||||
#SBATCH --cpus-per-gpu=10
|
||||
#SBATCH --mem=200G
|
||||
#SBATCH -t 24:00:00
|
||||
#SBATCH --output=logs/eval_gemma_matx_%j.log
|
||||
|
||||
set -uo pipefail
|
||||
SIZE=${1:?usage: <1k|2k> <benchmarks> <N> <tag>}
|
||||
BENCHES=${2:-gaia}
|
||||
N=${3:-2}
|
||||
TAG=${4:-smoke}
|
||||
PORT=8123
|
||||
|
||||
# Paths are resolved relative to the repo, or overridden by env:
|
||||
# OJ_ROOT repo checkout (default: two levels up from this script)
|
||||
# OJ_WORK working-state dir (default: $HOME/.openjarvis)
|
||||
# OJ_CKPT_ROOT trained SFT checkpoints (default: $OJ_WORK/lambda_ckpts)
|
||||
OJ_ROOT=${OJ_ROOT:-$(cd "$(dirname "${BASH_SOURCE[0]}")/../.." && pwd)}
|
||||
OJ_WORK=${OJ_WORK:-$HOME/.openjarvis}
|
||||
OJ_CKPT_ROOT=${OJ_CKPT_ROOT:-$OJ_WORK/lambda_ckpts}
|
||||
|
||||
cd "$OJ_ROOT"
|
||||
set -a
|
||||
source .env
|
||||
[ -f "$OJ_WORK/eval_key_override.sh" ] && source "$OJ_WORK/eval_key_override.sh"
|
||||
set +a
|
||||
export HF_HOME=${HF_HOME:?set HF_HOME to a cache dir with room for the weights}
|
||||
export PYTHONPATH="$PWD/src"
|
||||
# judge/expert cloud engine needs a config with an [engine.cloud] section
|
||||
export OPENJARVIS_CONFIG=${OPENJARVIS_CONFIG:?set OPENJARVIS_CONFIG to a cell config.toml}
|
||||
source .venv/bin/activate
|
||||
|
||||
if [ "$SIZE" = "base" ]; then
|
||||
CKPT=google/gemma-4-26B-A4B-it # un-fine-tuned base orchestrator (from HF cache)
|
||||
else
|
||||
CKPT=$OJ_CKPT_ROOT/sft_gemma_${SIZE}_bf16/epoch3
|
||||
fi
|
||||
OUT=results/orch-eval-gemma-${SIZE}-${TAG}
|
||||
mkdir -p "$OUT" logs
|
||||
echo "=== [$(date)] eval_gemma SIZE=$SIZE BENCHES=$BENCHES N=$N TAG=$TAG on $(hostname) ==="
|
||||
nvidia-smi --query-gpu=index,name,memory.total --format=csv,noheader
|
||||
|
||||
# ---- serve: TP=2 bf16 + MANDATORY no-NVLink L40S flags (else NCCL hangs) ----
|
||||
echo "=== [$(date)] launching vLLM (TP=2) serve of $CKPT ==="
|
||||
NCCL_P2P_DISABLE=1 NCCL_CUMEM_HOST_ENABLE=0 VLLM_WORKER_MULTIPROC_METHOD=spawn \
|
||||
PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True \
|
||||
python -m vllm.entrypoints.openai.api_server \
|
||||
--model "$CKPT" --served-model-name gemma-sft \
|
||||
--port $PORT --tensor-parallel-size 2 --dtype bfloat16 \
|
||||
--gpu-memory-utilization 0.90 --trust-remote-code \
|
||||
--max-model-len 32768 --enforce-eager --disable-custom-all-reduce \
|
||||
--enable-auto-tool-choice --tool-call-parser gemma4 \
|
||||
--max-num-batched-tokens 8192 --safetensors-load-strategy eager \
|
||||
--enable-prefix-caching > logs/serve_gemma_${SIZE}_${SLURM_JOB_ID}.log 2>&1 &
|
||||
SERVE_PID=$!
|
||||
echo "serve pid=$SERVE_PID log=logs/serve_gemma_${SIZE}_${SLURM_JOB_ID}.log"
|
||||
cleanup(){ echo "=== [$(date)] teardown kill $SERVE_PID ==="; kill $SERVE_PID 2>/dev/null; sleep 5; kill -9 $SERVE_PID 2>/dev/null; }
|
||||
trap cleanup EXIT
|
||||
|
||||
# ---- health poll (up to 25 min; TP + 52GB load is slow on L40S) ----
|
||||
echo "=== [$(date)] waiting for /v1/models ==="
|
||||
READY=0
|
||||
for i in $(seq 1 150); do
|
||||
if ! kill -0 $SERVE_PID 2>/dev/null; then echo "!! serve died during startup"; tail -40 logs/serve_gemma_${SIZE}_${SLURM_JOB_ID}.log; exit 3; fi
|
||||
if curl -sf http://localhost:$PORT/v1/models 2>/dev/null | grep -q gemma-sft; then READY=1; echo "=== [$(date)] READY after ${i}0s ==="; break; fi
|
||||
sleep 10
|
||||
done
|
||||
[ $READY -eq 1 ] || { echo "!! not ready in 25min"; tail -40 logs/serve_gemma_${SIZE}_${SLURM_JOB_ID}.log; exit 4; }
|
||||
|
||||
# ---- eval ----
|
||||
echo "=== [$(date)] running eval_orchestrator benches=$BENCHES n=$N ==="
|
||||
python scripts/orchestrator/eval_orchestrator.py \
|
||||
--benchmarks "$BENCHES" --n "$N" --seed 42 \
|
||||
--orchestrator-endpoint http://localhost:$PORT/v1 \
|
||||
--orchestrator-model gemma-sft --orchestrator-api-key EMPTY \
|
||||
--judge-model gemini-2.5-flash --judge-engine cloud \
|
||||
--max-workers 4 --max-turns 8 \
|
||||
--output-dir "$OUT"
|
||||
RC=$?
|
||||
echo "=== [$(date)] eval exit=$RC summary=$OUT/summary.json ==="
|
||||
echo "EVAL_GEMMA_DONE size=$SIZE tag=$TAG rc=$RC"
|
||||
exit $RC
|
||||
@@ -0,0 +1,25 @@
|
||||
compute_environment: LOCAL_MACHINE
|
||||
distributed_type: FSDP
|
||||
downcast_bf16: 'no'
|
||||
machine_rank: 0
|
||||
main_process_ip: null
|
||||
main_process_port: null
|
||||
main_training_function: main
|
||||
mixed_precision: bf16
|
||||
num_machines: 1
|
||||
num_processes: 4
|
||||
rdzv_backend: static
|
||||
same_network: true
|
||||
use_cpu: false
|
||||
fsdp_config:
|
||||
fsdp_version: 1
|
||||
fsdp_auto_wrap_policy: SIZE_BASED_WRAP
|
||||
fsdp_min_num_params: 100000000
|
||||
fsdp_sharding_strategy: FULL_SHARD
|
||||
fsdp_state_dict_type: FULL_STATE_DICT
|
||||
fsdp_offload_params: false
|
||||
fsdp_backward_prefetch: BACKWARD_PRE
|
||||
fsdp_forward_prefetch: false
|
||||
fsdp_use_orig_params: true
|
||||
fsdp_cpu_ram_efficient_loading: true
|
||||
fsdp_sync_module_states: true
|
||||
@@ -0,0 +1,25 @@
|
||||
compute_environment: LOCAL_MACHINE
|
||||
distributed_type: FSDP
|
||||
downcast_bf16: 'no'
|
||||
machine_rank: 0
|
||||
main_process_ip: null
|
||||
main_process_port: null
|
||||
main_training_function: main
|
||||
mixed_precision: bf16
|
||||
num_machines: 1
|
||||
num_processes: 8
|
||||
rdzv_backend: static
|
||||
same_network: true
|
||||
use_cpu: false
|
||||
fsdp_config:
|
||||
fsdp_version: 1
|
||||
fsdp_auto_wrap_policy: SIZE_BASED_WRAP
|
||||
fsdp_min_num_params: 100000000
|
||||
fsdp_sharding_strategy: FULL_SHARD
|
||||
fsdp_state_dict_type: FULL_STATE_DICT
|
||||
fsdp_offload_params: false
|
||||
fsdp_backward_prefetch: BACKWARD_PRE
|
||||
fsdp_forward_prefetch: false
|
||||
fsdp_use_orig_params: true
|
||||
fsdp_cpu_ram_efficient_loading: true
|
||||
fsdp_sync_module_states: true
|
||||
@@ -394,8 +394,8 @@ class LocalCloudAgent(BaseAgent):
|
||||
tools: Optional[list] = None,
|
||||
tool_choice: Optional[dict] = None,
|
||||
output_config: Optional[dict] = None,
|
||||
timeout: float = 600.0,
|
||||
max_retries: int = 12,
|
||||
timeout: float = 60.0,
|
||||
max_retries: int = 2,
|
||||
trace_role: str = "cloud",
|
||||
) -> Tuple[str, int, int, int]:
|
||||
"""Single Anthropic call. Returns (text, p_tok, c_tok, n_web_searches).
|
||||
@@ -471,7 +471,7 @@ class LocalCloudAgent(BaseAgent):
|
||||
response_format: Optional[dict] = None,
|
||||
tools: Optional[list] = None,
|
||||
tool_choice: Optional[Any] = None,
|
||||
timeout: float = 600.0,
|
||||
timeout: float = 60.0,
|
||||
trace_role: str = "cloud",
|
||||
) -> Tuple[str, int, int]:
|
||||
"""Single OpenAI call. Returns (text, p_tok, c_tok). Trace-captured;
|
||||
@@ -537,7 +537,7 @@ class LocalCloudAgent(BaseAgent):
|
||||
system: Optional[str] = None,
|
||||
max_tokens: int = 4096,
|
||||
temperature: float = 0.0,
|
||||
timeout: float = 600.0,
|
||||
timeout: float = 60.0,
|
||||
trace_role: str = "cloud",
|
||||
extra_body: Optional[Dict[str, Any]] = None,
|
||||
) -> Tuple[str, int, int]:
|
||||
|
||||
@@ -11,10 +11,12 @@ from __future__ import annotations
|
||||
|
||||
# USD per million tokens, (input, output). Local models = 0.
|
||||
PRICES: dict[str, tuple[float, float]] = {
|
||||
"claude-opus-4-8": (5.00, 25.0),
|
||||
"claude-opus-4-7": (5.00, 25.0),
|
||||
"claude-sonnet-4-6": (3.00, 15.0),
|
||||
"claude-haiku-4-5": (1.00, 5.00),
|
||||
"claude-haiku-4-5-20251001": (1.00, 5.00),
|
||||
"gpt-5.5": (2.00, 15.0),
|
||||
"gpt-5": (1.25, 10.0),
|
||||
"gpt-5-mini": (0.25, 2.00),
|
||||
"gpt-5-mini-2025-08-07": (0.25, 2.00),
|
||||
@@ -30,6 +32,14 @@ PRICES: dict[str, tuple[float, float]] = {
|
||||
"qwen/qwen-2.5-coder-32b-instruct": (0.08, 0.18),
|
||||
"qwen/qwen3-32b": (0.10, 0.30),
|
||||
"meta-llama/llama-3.3-70b-instruct": (0.13, 0.39),
|
||||
# OpenRouter slugs for the orchestrator's local-OSS class when routed via
|
||||
# OpenRouter instead of self-hosted vLLM. ESTIMATES — no public list price
|
||||
# exists yet for these Qwen3.5/3.6 builds; scaled by active-param size.
|
||||
# VERIFY against openrouter.ai before trusting the cost-aware reward numbers.
|
||||
"qwen/qwen3.5-9b": (0.05, 0.10),
|
||||
"qwen/qwen3.6-27b": (0.10, 0.30),
|
||||
"qwen/qwen3.5-122b-a10b": (0.20, 0.60),
|
||||
"qwen/qwen3.5-397b-a17b": (0.40, 1.20),
|
||||
}
|
||||
|
||||
# Models whose API rejects an explicit `temperature` param — callers should
|
||||
@@ -38,6 +48,10 @@ NO_TEMP_PREFIXES: tuple[str, ...] = (
|
||||
"claude-opus-4-7",
|
||||
"claude-sonnet-4-7",
|
||||
"claude-haiku-4-7",
|
||||
# 4-8 family also rejects `temperature` ("deprecated for this model").
|
||||
"claude-opus-4-8",
|
||||
"claude-sonnet-4-8",
|
||||
"claude-haiku-4-8",
|
||||
)
|
||||
|
||||
|
||||
|
||||
@@ -0,0 +1,783 @@
|
||||
"""Faithful ToolOrchestra "unified tool calling" registry (arXiv:2511.21689 §3.1).
|
||||
|
||||
The paper exposes **every tool AND every model through a single flat tool
|
||||
interface** — each is its own named function with a description and a typed
|
||||
parameter schema, and for each training instance a *random subset* of tools is
|
||||
sampled with *randomized pricing* (§3.3, "General tool configuration"). This is
|
||||
unlike the eval-port shortcut in ``toolorchestra.py``, which collapses the whole
|
||||
catalog into three meta-tools (``search``/``enhance_reasoning``/``answer``) with
|
||||
a ``model`` slot. This module restores the faithful design.
|
||||
|
||||
Each :class:`ExpertTool` knows:
|
||||
|
||||
* the orchestrator-visible ``name`` / ``description`` / param schema (what goes
|
||||
into the tools JSON the policy conditions on), and
|
||||
* the concrete backend (``backend_type`` + ``model`` + ``base_url``) so a caller
|
||||
can turn it into the worker dict that ``toolorchestra._call_worker`` dispatches.
|
||||
|
||||
Everything here is pure data + deterministic transforms (no network, no model
|
||||
calls), so the spec building, sampling, and pricing logic is offline-testable.
|
||||
Dispatch stays in ``toolorchestra.py`` (via :func:`to_worker_dict`) to avoid a
|
||||
circular import.
|
||||
"""
|
||||
|
||||
from __future__ import annotations
|
||||
|
||||
import os
|
||||
import random
|
||||
import re
|
||||
from dataclasses import dataclass, field
|
||||
from typing import Dict, List, Optional
|
||||
|
||||
from openjarvis.agents.hybrid._prices import PRICES
|
||||
|
||||
# Kinds of tool in the unified interface.
|
||||
KIND_MODEL = "model" # an LLM exposed as a tool (the paper's "models as tools")
|
||||
KIND_WEB_SEARCH = "web_search"
|
||||
KIND_LOCAL_SEARCH = "local_search"
|
||||
KIND_CODE = "code_interpreter"
|
||||
KIND_TOOL = "tool" # a bridged real OpenJarvis tool (custom param schema)
|
||||
|
||||
VALID_KINDS = (KIND_MODEL, KIND_WEB_SEARCH, KIND_LOCAL_SEARCH, KIND_CODE, KIND_TOOL)
|
||||
|
||||
# Flat-catalog category label, surfaced in each tool's spec so the orchestrator
|
||||
# can tell tool types apart without us imposing any hierarchy (the menu stays
|
||||
# flat — this is just a tag).
|
||||
CATEGORY_BASIC = "basic_tool"
|
||||
# Two-model-class taxonomy for the orchestrator catalog — the only model tiers.
|
||||
# (The legacy default_catalog still uses CATEGORY_GENERALIST/SPECIALIZED below.)
|
||||
CATEGORY_CLOUD_FRONTIER = "cloud_frontier"
|
||||
CATEGORY_LOCAL_OSS = "local_open_source"
|
||||
# Legacy tiers kept only for default_catalog (build_unified_sft.py); not used by
|
||||
# orchestrator_catalog. Superseded by the two-class taxonomy above.
|
||||
CATEGORY_GENERALIST = "generalist_model"
|
||||
CATEGORY_SPECIALIZED = "specialized_model"
|
||||
|
||||
# Backend dispatch types understood by ``toolorchestra._call_worker`` (plus the
|
||||
# ``openjarvis-tool`` bridge, dispatched in ``unified.make_dispatch`` via the
|
||||
# OpenJarvis ToolExecutor rather than ``_call_worker``).
|
||||
VALID_BACKENDS = (
|
||||
"vllm", "openai", "anthropic", "gemini", "openrouter",
|
||||
"anthropic-web-search", "tavily-search", "modal-python",
|
||||
"openjarvis-tool",
|
||||
)
|
||||
|
||||
|
||||
@dataclass(frozen=True)
|
||||
class ExpertTool:
|
||||
"""One entry in the unified tool catalog.
|
||||
|
||||
``price_in`` / ``price_out`` are USD per 1M tokens (0.0 for local / non-LLM
|
||||
tools). ``latency_s`` is a rough average used only to populate the
|
||||
description's cost/latency line — the orchestrator was trained to read that
|
||||
table, so we surface it verbatim in the spec.
|
||||
"""
|
||||
|
||||
name: str
|
||||
kind: str
|
||||
backend_type: str
|
||||
summary: str
|
||||
model: Optional[str] = None
|
||||
base_url: Optional[str] = None
|
||||
price_in: float = 0.0
|
||||
price_out: float = 0.0
|
||||
latency_s: float = 5.0
|
||||
category: str = "" # cloud_frontier | local_open_source | basic_tool
|
||||
# Optional explicit JSON-schema for the tool's arguments. Set for bridged
|
||||
# real OpenJarvis tools (``openjarvis-tool`` backend) whose params don't fit
|
||||
# the fixed kind-based schemas; takes precedence over the kind default.
|
||||
param_schema: Optional[dict] = None
|
||||
# When True, ``description()`` omits the price/latency line — used by the
|
||||
# anonymized catalog so the orchestrator can't route on cost/identity.
|
||||
hide_cost: bool = False
|
||||
|
||||
def __post_init__(self) -> None:
|
||||
if self.kind not in VALID_KINDS:
|
||||
raise ValueError(f"{self.name}: invalid kind {self.kind!r}")
|
||||
if self.backend_type not in VALID_BACKENDS:
|
||||
raise ValueError(f"{self.name}: invalid backend {self.backend_type!r}")
|
||||
if self.kind == KIND_MODEL and not self.model:
|
||||
raise ValueError(f"{self.name}: model-kind tool needs a concrete model")
|
||||
|
||||
# ---- orchestrator-visible spec -------------------------------------
|
||||
|
||||
def _param_schema(self) -> Dict[str, object]:
|
||||
"""JSON-schema for the tool's arguments (one typed param per kind).
|
||||
|
||||
An explicit ``param_schema`` (set by :func:`openjarvis_tool` for bridged
|
||||
real tools) overrides the kind-based default.
|
||||
"""
|
||||
if self.param_schema is not None:
|
||||
return self.param_schema
|
||||
if self.kind == KIND_WEB_SEARCH or self.kind == KIND_LOCAL_SEARCH:
|
||||
return {
|
||||
"type": "object",
|
||||
"properties": {
|
||||
"query": {
|
||||
"type": "string",
|
||||
"description": "Search query string.",
|
||||
}
|
||||
},
|
||||
"required": ["query"],
|
||||
}
|
||||
if self.kind == KIND_CODE:
|
||||
return {
|
||||
"type": "object",
|
||||
"properties": {
|
||||
"code": {
|
||||
"type": "string",
|
||||
"description": "Python code to execute. Print results.",
|
||||
}
|
||||
},
|
||||
"required": ["code"],
|
||||
}
|
||||
# model tool
|
||||
return {
|
||||
"type": "object",
|
||||
"properties": {
|
||||
"input": {
|
||||
"type": "string",
|
||||
"description": "The sub-question or instruction for this model.",
|
||||
}
|
||||
},
|
||||
"required": ["input"],
|
||||
}
|
||||
|
||||
def description(self) -> str:
|
||||
"""Full description incl. the price/latency line (paper bakes this in)."""
|
||||
if self.hide_cost:
|
||||
return self.summary
|
||||
if self.kind == KIND_MODEL:
|
||||
cost_line = (
|
||||
f" Pricing: ${self.price_in:.2f}/1M input, "
|
||||
f"${self.price_out:.2f}/1M output; avg latency ~{self.latency_s:.0f}s."
|
||||
)
|
||||
else:
|
||||
cost_line = f" Avg latency ~{self.latency_s:.0f}s."
|
||||
return self.summary.rstrip(".") + "." + cost_line
|
||||
|
||||
def to_spec(self) -> Dict[str, object]:
|
||||
"""OpenAI-style tool spec the orchestrator conditions on.
|
||||
|
||||
Flat list, but each function carries a ``category`` tag so the policy can
|
||||
distinguish generalist vs specialized models vs basic tools.
|
||||
"""
|
||||
fn: Dict[str, object] = {
|
||||
"name": self.name,
|
||||
"description": self.description(),
|
||||
"parameters": self._param_schema(),
|
||||
}
|
||||
if self.category:
|
||||
fn["category"] = self.category
|
||||
return {"type": "function", "function": fn}
|
||||
|
||||
|
||||
def _price(model: str) -> tuple[float, float]:
|
||||
return PRICES.get(model, (0.0, 0.0))
|
||||
|
||||
|
||||
def _tool_name(model: str) -> str:
|
||||
"""Tool-safe function name derived from a model id (``qwen3-8b`` -> ``qwen3_8b``).
|
||||
|
||||
Strips any provider prefix and replaces non-alphanumerics with underscores so
|
||||
the catalog exposes one named tool per concrete model.
|
||||
"""
|
||||
base = model.split("/")[-1].lower()
|
||||
safe = re.sub(r"[^a-z0-9]+", "_", base).strip("_")
|
||||
return safe or "local_model"
|
||||
|
||||
|
||||
_ANON_ALPHABET = "abcdefghijklmnopqrstuvwxyz0123456789"
|
||||
|
||||
# Rough size/scale hint per model — a capability signal WITHOUT the brand name,
|
||||
# so the orchestrator can route big-vs-small on task difficulty but can't route on
|
||||
# a proprietary name/tier prior. Closed models have no public param count, so they
|
||||
# get a scale tier; open models get their param count (else parsed from the name).
|
||||
_MODEL_SIZE = {
|
||||
# The real orchestrator_catalog: 2 cloud frontier + 4 local OSS Qwen.
|
||||
"gpt-5.5": "frontier-scale",
|
||||
"claude-opus-4-8": "frontier-scale",
|
||||
"Qwen/Qwen3.5-9B": "~9B params",
|
||||
"Qwen/Qwen3.6-27B-FP8": "~27B params",
|
||||
"Qwen/Qwen3.5-122B-A10B-FP8": "~122B total, ~10B active (MoE)",
|
||||
"Qwen/Qwen3.5-397B-A17B-FP8": "~397B total, ~17B active (MoE)",
|
||||
# OpenRouter slugs (the tool.model id when these route through OpenRouter).
|
||||
"qwen/qwen3.5-122b-a10b": "~122B total, ~10B active (MoE)",
|
||||
"qwen/qwen3.5-397b-a17b": "~397B total, ~17B active (MoE)",
|
||||
}
|
||||
|
||||
|
||||
def _size_hint(model: str) -> str:
|
||||
"""Anonymized scale descriptor for an expert (no brand). Falls back to a
|
||||
param count parsed from the model id (``...-9b`` -> ``~9B params``)."""
|
||||
if model in _MODEL_SIZE:
|
||||
return _MODEL_SIZE[model]
|
||||
m = re.search(r"(\d+(?:\.\d+)?)\s*b\b", model.lower())
|
||||
return f"~{m.group(1)}B params" if m else "unspecified scale"
|
||||
|
||||
|
||||
def _class_hint(tool) -> str:
|
||||
"""Which of the two model classes this expert belongs to (anonymized)."""
|
||||
cat = getattr(tool, "category", "")
|
||||
if cat == CATEGORY_CLOUD_FRONTIER:
|
||||
return "cloud frontier"
|
||||
if cat == CATEGORY_LOCAL_OSS:
|
||||
return "local open-source"
|
||||
return ""
|
||||
|
||||
|
||||
def _cost_tier_hint(tool) -> str:
|
||||
"""Coarse cost tier (not exact pricing — that's the bias we anonymize away).
|
||||
Shown so the policy learns the strong-but-expensive vs cheap tradeoff instead
|
||||
of always delegating to the biggest model. Driven by class: cloud frontier is
|
||||
expensive, self-hosted open-source is cheap."""
|
||||
cat = getattr(tool, "category", "")
|
||||
if cat == CATEGORY_CLOUD_FRONTIER:
|
||||
return "expensive"
|
||||
if cat == CATEGORY_LOCAL_OSS:
|
||||
return "cheap"
|
||||
if getattr(tool, "backend_type", "") == "vllm" or (tool.price_out or 0) <= 0:
|
||||
return "cheap"
|
||||
po = tool.price_out
|
||||
return "cheap" if po < 5 else "moderate" if po < 20 else "expensive"
|
||||
|
||||
|
||||
def anonymize_tools(tools, rng):
|
||||
"""Strip model identity for unbiased routing data.
|
||||
|
||||
Each MODEL expert is replaced with an opaque random label
|
||||
(``expert_<4 rand>``), a uniform description, a uniform category and no
|
||||
price/latency line — so the orchestrator cannot route on a model's name,
|
||||
position, cost, or tier (all of which we found dominate the choice). The full
|
||||
list is shuffled to kill position bias. Basic tools (calculator, web_search,
|
||||
…) keep their real names — the policy must still know what they do.
|
||||
|
||||
Returns ``(anon_tools, anon_to_real)`` where ``anon_to_real`` maps each opaque
|
||||
label back to the real tool name (for offline analysis only; never shown to
|
||||
the model). ``.model`` is preserved on each tool so dispatch still reaches the
|
||||
right backend. Pass a fresh ``rng`` per rollout so labels don't stabilise.
|
||||
"""
|
||||
from dataclasses import replace
|
||||
|
||||
anon_to_real: Dict[str, str] = {}
|
||||
experts: List[ExpertTool] = []
|
||||
basics: List[ExpertTool] = []
|
||||
for t in tools:
|
||||
if t.kind == KIND_MODEL:
|
||||
tag = "model_" + "".join(rng.choice(_ANON_ALPHABET) for _ in range(4))
|
||||
while tag in anon_to_real:
|
||||
tag = "model_" + "".join(rng.choice(_ANON_ALPHABET) for _ in range(4))
|
||||
anon_to_real[tag] = t.name
|
||||
bits = [b for b in (_class_hint(t), _size_hint(t.model), _cost_tier_hint(t)) if b]
|
||||
experts.append(replace(
|
||||
t, name=tag,
|
||||
summary="Another model — " + ", ".join(bits) + ". Send it a sub-question.",
|
||||
category="model", hide_cost=True,
|
||||
))
|
||||
else:
|
||||
basics.append(t)
|
||||
# Shuffle WITHIN the experts to kill per-expert position bias, but keep all
|
||||
# experts as a block on top and the basic tools underneath — a clean split
|
||||
# (experts first) rather than experts and utilities interleaved.
|
||||
rng.shuffle(experts)
|
||||
out = experts + basics
|
||||
return out, anon_to_real
|
||||
|
||||
|
||||
# Default catalog: the paper's tool categories, mapped onto the models/tools
|
||||
# OpenJarvis can actually call. One named tool per model (faithful §3.1).
|
||||
def default_catalog(
|
||||
*,
|
||||
local_model: Optional[str] = None,
|
||||
local_endpoint: Optional[str] = None,
|
||||
) -> List[ExpertTool]:
|
||||
"""Return the full unified tool catalog.
|
||||
|
||||
``local_model`` / ``local_endpoint`` wire the on-device vLLM tool when a
|
||||
local backbone is served; omitted → the local model tool is left out.
|
||||
"""
|
||||
cat: List[ExpertTool] = []
|
||||
|
||||
# ---- generalist / frontier models (one named tool per model VERSION) ----
|
||||
for model, summary, lat in [
|
||||
("gpt-5",
|
||||
"Frontier generalist (GPT-5). Strongest reasoning across domains.", 30.0),
|
||||
("gpt-5-mini",
|
||||
"Mid-tier generalist (GPT-5-mini). Solid reasoning, much cheaper.", 15.0),
|
||||
("gpt-4o",
|
||||
"Fast generalist (GPT-4o). Good for simple steps and formatting.", 8.0),
|
||||
("claude-opus-4-7",
|
||||
"Frontier generalist (Claude Opus 4.7). Strong long-horizon reasoning.", 26.0),
|
||||
("claude-sonnet-4-6",
|
||||
"Strong generalist (Claude Sonnet 4.6). Balanced cost/capability.", 15.0),
|
||||
("gemini-2.5-pro",
|
||||
"Frontier generalist (Gemini 2.5 Pro). Strong multimodal reasoning.", 20.0),
|
||||
("gemini-2.5-flash",
|
||||
"Cheap fast generalist (Gemini 2.5 Flash).", 8.0),
|
||||
("meta-llama/llama-3.3-70b-instruct",
|
||||
"Open generalist (Llama-3.3-70B). Decent general knowledge, low cost.", 10.0),
|
||||
("qwen/qwen3-32b",
|
||||
"Open generalist (Qwen3-32B). Strong math/science reasoning, low cost.", 9.0),
|
||||
]:
|
||||
ep = ("openai" if model.startswith("gpt") else
|
||||
"anthropic" if model.startswith("claude") else
|
||||
"gemini" if model.startswith("gemini") else "openrouter")
|
||||
pi, po = _price(model)
|
||||
cat.append(ExpertTool(
|
||||
name=_tool_name(model), kind=KIND_MODEL, backend_type=ep, summary=summary,
|
||||
model=model, price_in=pi, price_out=po, latency_s=lat,
|
||||
category=CATEGORY_GENERALIST,
|
||||
))
|
||||
|
||||
# ---- specialized: code ----
|
||||
coder = "qwen/qwen-2.5-coder-32b-instruct"
|
||||
pi, po = _price(coder)
|
||||
cat.append(ExpertTool(
|
||||
name=_tool_name(coder), kind=KIND_MODEL, backend_type="openrouter",
|
||||
summary="Specialized code model (Qwen2.5-Coder-32B). Writes/debugs code.",
|
||||
model=coder, price_in=pi, price_out=po, latency_s=9.0,
|
||||
category=CATEGORY_SPECIALIZED,
|
||||
))
|
||||
|
||||
# ---- local backbone as a tool (on-device vLLM), if served ----
|
||||
# Named after the actual served model (faithful "one named tool per model"),
|
||||
# not a generic "local_model" — e.g. "qwen3-8b" -> tool "qwen3_8b".
|
||||
if local_model and local_endpoint:
|
||||
cat.append(ExpertTool(
|
||||
name=_tool_name(local_model), kind=KIND_MODEL, backend_type="vllm",
|
||||
summary=(f"On-device open model ({local_model}) served locally. Cheap "
|
||||
"and private; good for extraction, formatting, arithmetic on "
|
||||
"given data."),
|
||||
model=local_model, base_url=local_endpoint,
|
||||
price_in=0.0, price_out=0.0, latency_s=2.0,
|
||||
category=CATEGORY_GENERALIST,
|
||||
))
|
||||
|
||||
# ---- basic tools ----
|
||||
cat.append(ExpertTool(
|
||||
name="web_search", kind=KIND_WEB_SEARCH, backend_type="tavily-search",
|
||||
summary="Web search (Tavily). Use for facts that need a live lookup.",
|
||||
model="tavily", latency_s=8.0, category=CATEGORY_BASIC,
|
||||
))
|
||||
cat.append(ExpertTool(
|
||||
name="code_interpreter", kind=KIND_CODE, backend_type="modal-python",
|
||||
summary="Python sandbox. Execute code and return stdout/stderr.",
|
||||
model="modal-python", latency_s=6.0, category=CATEGORY_BASIC,
|
||||
))
|
||||
|
||||
return cat
|
||||
|
||||
|
||||
def openjarvis_tool(
|
||||
registered_name: str,
|
||||
*,
|
||||
summary: str,
|
||||
params: dict,
|
||||
latency_s: float = 5.0,
|
||||
) -> ExpertTool:
|
||||
"""Build an :class:`ExpertTool` that bridges a real OpenJarvis tool.
|
||||
|
||||
``registered_name`` is the tool's key in ``ToolRegistry`` (e.g. ``calculator``,
|
||||
``shell_exec``). ``params`` is the JSON-schema *properties*-style dict for the
|
||||
tool's arguments; it is surfaced verbatim by :meth:`ExpertTool.to_spec`. The
|
||||
resulting tool dispatches through the OpenJarvis ``ToolExecutor`` (backend
|
||||
``openjarvis-tool``) rather than ``_call_worker``.
|
||||
"""
|
||||
return ExpertTool(
|
||||
name=registered_name,
|
||||
kind=KIND_TOOL,
|
||||
backend_type="openjarvis-tool",
|
||||
summary=summary,
|
||||
model=registered_name,
|
||||
latency_s=latency_s,
|
||||
category=CATEGORY_BASIC,
|
||||
param_schema=params,
|
||||
)
|
||||
|
||||
|
||||
# Real OpenJarvis tools bridged into the orchestrator catalog as basic tools.
|
||||
# Names must match the ``ToolRegistry`` keys (confirmed present: calculator,
|
||||
# shell_exec, file_read, file_write, http_request).
|
||||
def _openjarvis_basic_tools() -> List[ExpertTool]:
|
||||
def obj(properties: dict, required: List[str]) -> dict:
|
||||
return {"type": "object", "properties": properties, "required": required}
|
||||
|
||||
return [
|
||||
openjarvis_tool(
|
||||
"calculator",
|
||||
summary="Evaluate an arithmetic / math expression and return the result.",
|
||||
params=obj(
|
||||
{"expression": {"type": "string",
|
||||
"description": "Math expression to evaluate."}},
|
||||
["expression"],
|
||||
),
|
||||
latency_s=1.0,
|
||||
),
|
||||
openjarvis_tool(
|
||||
"shell_exec",
|
||||
summary=("Run a shell command and return its stdout/stderr. Critical "
|
||||
"for terminal / TerminalBench-style tasks."),
|
||||
params=obj(
|
||||
{"command": {"type": "string",
|
||||
"description": "Shell command to execute."}},
|
||||
["command"],
|
||||
),
|
||||
latency_s=4.0,
|
||||
),
|
||||
openjarvis_tool(
|
||||
"file_read",
|
||||
summary="Read the contents of a file at the given path.",
|
||||
params=obj(
|
||||
{"path": {"type": "string", "description": "Path of the file to read."}},
|
||||
["path"],
|
||||
),
|
||||
latency_s=1.0,
|
||||
),
|
||||
openjarvis_tool(
|
||||
"file_write",
|
||||
summary="Write content to a file at the given path.",
|
||||
params=obj(
|
||||
{"path": {"type": "string", "description": "Path of the file to write."},
|
||||
"content": {"type": "string", "description": "Content to write."}},
|
||||
["path", "content"],
|
||||
),
|
||||
latency_s=1.0,
|
||||
),
|
||||
openjarvis_tool(
|
||||
"http_request",
|
||||
summary="Make an HTTP request to a URL and return the response body.",
|
||||
params={
|
||||
"type": "object",
|
||||
"properties": {
|
||||
"url": {"type": "string", "description": "Request URL."},
|
||||
"method": {"type": "string",
|
||||
"description": "HTTP method (GET, POST, ...). Default GET."},
|
||||
},
|
||||
"required": ["url"],
|
||||
},
|
||||
latency_s=4.0,
|
||||
),
|
||||
openjarvis_tool(
|
||||
"think",
|
||||
summary=("Record a private reasoning step (scratchpad). No external "
|
||||
"effect; use to plan before acting on hard reasoning tasks."),
|
||||
params=obj(
|
||||
{"thought": {"type": "string",
|
||||
"description": "Your reasoning or thought process."}},
|
||||
["thought"],
|
||||
),
|
||||
latency_s=0.5,
|
||||
),
|
||||
openjarvis_tool(
|
||||
"apply_patch",
|
||||
summary=("Apply a unified-diff patch to a file. Use to edit code for "
|
||||
"terminal / SWE-style tasks."),
|
||||
params=obj(
|
||||
{"patch": {"type": "string",
|
||||
"description": "The unified diff patch text to apply."},
|
||||
"path": {"type": "string",
|
||||
"description": "Target file path (auto-detected from the "
|
||||
"patch header if omitted)."}},
|
||||
["patch"],
|
||||
),
|
||||
latency_s=2.0,
|
||||
),
|
||||
openjarvis_tool(
|
||||
"pdf_extract",
|
||||
summary=("Extract text from a PDF file. Use for GAIA-style tasks with "
|
||||
"PDF attachments."),
|
||||
params=obj(
|
||||
{"file_path": {"type": "string",
|
||||
"description": "Path to the PDF file."},
|
||||
"pages": {"type": "string",
|
||||
"description": "Page range, e.g. '1-5' or '1,3,5'. "
|
||||
"Omit for all pages."}},
|
||||
["file_path"],
|
||||
),
|
||||
latency_s=3.0,
|
||||
),
|
||||
openjarvis_tool(
|
||||
"db_query",
|
||||
summary=("Run a SQL query against a SQLite/Postgres database and return "
|
||||
"rows. Read-only by default."),
|
||||
params=obj(
|
||||
{"query": {"type": "string",
|
||||
"description": "SQL query to execute."},
|
||||
"db_path": {"type": "string",
|
||||
"description": "Path to a SQLite DB file. Defaults to "
|
||||
"in-memory."},
|
||||
"read_only": {"type": "boolean",
|
||||
"description": "Restrict to SELECT/EXPLAIN/PRAGMA. "
|
||||
"Default: true."}},
|
||||
["query"],
|
||||
),
|
||||
latency_s=3.0,
|
||||
),
|
||||
]
|
||||
|
||||
|
||||
# Local-Cloud Hybrid orchestrator catalog. The menu is two model classes — cloud
|
||||
# frontier models and open-source models — plus the basic tools (web search, code
|
||||
# interpreter, and the bridged real OpenJarvis tools). The orchestrator model
|
||||
# itself is NOT in the catalog.
|
||||
#
|
||||
# Routing default is OpenRouter for every model tool, so the catalog works with
|
||||
# no self-hosted servers. A model is routed to local vLLM instead when its id is
|
||||
# in ``local_endpoints`` or ``model_backends`` overrides it. The
|
||||
# orchestrator-visible tool *name* is always derived from the canonical id, so
|
||||
# routing can change (vLLM <-> OpenRouter <-> native API) without shifting the
|
||||
# menu the policy was trained on.
|
||||
_LOCAL_OSS_MODELS = (
|
||||
# (canonical_id, openrouter_slug)
|
||||
("Qwen/Qwen3.5-9B", "qwen/qwen3.5-9b"),
|
||||
("Qwen/Qwen3.6-27B-FP8", "qwen/qwen3.6-27b"),
|
||||
("Qwen/Qwen3.5-122B-A10B-FP8", "qwen/qwen3.5-122b-a10b"),
|
||||
("Qwen/Qwen3.5-397B-A17B-FP8", "qwen/qwen3.5-397b-a17b"),
|
||||
)
|
||||
|
||||
_CLOUD_FRONTIER_MODELS = (
|
||||
# (canonical, native_backend, openrouter_slug, summary, latency_s)
|
||||
# Neutral, uniform summaries (no capability ranking) so the orchestrator
|
||||
# doesn't just pick whichever model is labelled "strongest" — routing should
|
||||
# be learned from the reward, not hand-labelled here.
|
||||
("gpt-5.5", "openai", "openai/gpt-5.5",
|
||||
"Expert model (GPT-5.5).", 30.0),
|
||||
("claude-opus-4-8", "anthropic", "anthropic/claude-opus-4.8",
|
||||
"Expert model (Claude Opus 4.8).", 26.0),
|
||||
)
|
||||
|
||||
|
||||
def _model_tool(
|
||||
canonical: str,
|
||||
*,
|
||||
native_backend: str,
|
||||
or_slug: str,
|
||||
summary: str,
|
||||
lat: float,
|
||||
category: str,
|
||||
local_endpoints: Dict[str, str],
|
||||
model_backends: Dict[str, str],
|
||||
openrouter_slugs: Dict[str, str],
|
||||
) -> ExpertTool:
|
||||
"""Build one model tool, resolving its backend.
|
||||
|
||||
Backend precedence: explicit ``model_backends[canonical]`` > vLLM if the model
|
||||
has a ``local_endpoints`` entry > the model's NATIVE provider (openai /
|
||||
anthropic / gemini) when it has one > OpenRouter. Frontier models thus hit
|
||||
their first-party API by default (OpenRouter's gpt-5.5 was returning 400s);
|
||||
OSS Qwen experts (native_backend="vllm") fall through to OpenRouter.
|
||||
"""
|
||||
backend = (
|
||||
model_backends.get(canonical)
|
||||
or ("vllm" if canonical in local_endpoints else None)
|
||||
or (native_backend if native_backend in ("openai", "anthropic", "gemini")
|
||||
else "openrouter")
|
||||
)
|
||||
name = _tool_name(canonical)
|
||||
if backend == "vllm":
|
||||
# Self-hosted: free per the cost model.
|
||||
return ExpertTool(
|
||||
name=name, kind=KIND_MODEL, backend_type="vllm", summary=summary,
|
||||
model=canonical, base_url=local_endpoints.get(canonical),
|
||||
price_in=0.0, price_out=0.0, latency_s=lat, category=category,
|
||||
)
|
||||
if backend == "openrouter":
|
||||
slug = openrouter_slugs.get(canonical, or_slug)
|
||||
pi, po = _price(slug)
|
||||
if (pi, po) == (0.0, 0.0): # fall back to the canonical id's price
|
||||
pi, po = _price(canonical)
|
||||
return ExpertTool(
|
||||
name=name, kind=KIND_MODEL, backend_type="openrouter", summary=summary,
|
||||
model=slug, base_url=None, price_in=pi, price_out=po,
|
||||
latency_s=lat, category=category,
|
||||
)
|
||||
# native provider API (openai / anthropic / gemini)
|
||||
pi, po = _price(canonical)
|
||||
return ExpertTool(
|
||||
name=name, kind=KIND_MODEL, backend_type=backend, summary=summary,
|
||||
model=canonical, price_in=pi, price_out=po, latency_s=lat, category=category,
|
||||
)
|
||||
|
||||
|
||||
def orchestrator_catalog(
|
||||
*,
|
||||
local_endpoints: Optional[Dict[str, str]] = None,
|
||||
model_backends: Optional[Dict[str, str]] = None,
|
||||
openrouter_slugs: Optional[Dict[str, str]] = None,
|
||||
include_tools: bool = True,
|
||||
) -> List[ExpertTool]:
|
||||
"""Return the orchestrator's tool catalog: two model classes + basic tools.
|
||||
|
||||
Routing for the model tools defaults to **OpenRouter** (so the catalog works
|
||||
with no self-hosted servers). Overrides, in precedence order:
|
||||
|
||||
* ``model_backends`` maps a canonical model id -> ``"vllm" | "openrouter" |
|
||||
"openai" | "anthropic" | "gemini"`` to force that model's backend.
|
||||
* ``local_endpoints`` maps a canonical id (e.g. ``"Qwen/Qwen3.5-9B"``) to a
|
||||
vLLM ``base_url``; a model present here routes to vLLM (free) unless
|
||||
``model_backends`` says otherwise.
|
||||
* ``openrouter_slugs`` overrides the per-model OpenRouter slug used when a
|
||||
model routes through OpenRouter.
|
||||
|
||||
``include_tools`` (default True) appends the basic tools — web search, code
|
||||
interpreter, and the bridged real OpenJarvis tools (calculator, shell_exec,
|
||||
file_read, file_write, http_request).
|
||||
"""
|
||||
local_endpoints = local_endpoints or {}
|
||||
model_backends = model_backends or {}
|
||||
openrouter_slugs = openrouter_slugs or {}
|
||||
cat: List[ExpertTool] = []
|
||||
|
||||
# Env-gated expert exclusion: OJ_EXCLUDE_EXPERTS is a comma-separated list of
|
||||
# case-insensitive substrings matched against a model's canonical id. Any
|
||||
# match is skipped from the catalog. Used to temporarily drop unreliable
|
||||
# experts (e.g. OpenRouter giants during a provider outage) without editing
|
||||
# the registry — unset the var to restore them.
|
||||
_excl = {s.strip().lower() for s in os.environ.get("OJ_EXCLUDE_EXPERTS", "").split(",") if s.strip()}
|
||||
def _excluded(canonical: str) -> bool:
|
||||
c = canonical.lower()
|
||||
return any(x in c for x in _excl)
|
||||
|
||||
# ---- cloud frontier models ----
|
||||
for canonical, native_backend, or_slug, summary, lat in _CLOUD_FRONTIER_MODELS:
|
||||
if _excluded(canonical):
|
||||
continue
|
||||
cat.append(_model_tool(
|
||||
canonical, native_backend=native_backend, or_slug=or_slug,
|
||||
summary=summary, lat=lat, category=CATEGORY_CLOUD_FRONTIER,
|
||||
local_endpoints=local_endpoints, model_backends=model_backends,
|
||||
openrouter_slugs=openrouter_slugs,
|
||||
))
|
||||
|
||||
# ---- open-source models (OpenRouter by default; vLLM when an endpoint or
|
||||
# a model_backends override is supplied) ----
|
||||
for canonical, or_slug in _LOCAL_OSS_MODELS:
|
||||
if _excluded(canonical):
|
||||
continue
|
||||
cat.append(_model_tool(
|
||||
canonical, native_backend="vllm", or_slug=or_slug,
|
||||
summary=f"Expert model ({canonical}).",
|
||||
lat=4.0, category=CATEGORY_LOCAL_OSS,
|
||||
local_endpoints=local_endpoints, model_backends=model_backends,
|
||||
openrouter_slugs=openrouter_slugs,
|
||||
))
|
||||
|
||||
if include_tools:
|
||||
# ---- basic tools ----
|
||||
cat.append(ExpertTool(
|
||||
name="web_search", kind=KIND_WEB_SEARCH, backend_type="tavily-search",
|
||||
summary="Web search (Tavily). Use for facts that need a live lookup.",
|
||||
model="tavily", latency_s=8.0, category=CATEGORY_BASIC,
|
||||
))
|
||||
cat.append(ExpertTool(
|
||||
name="code_interpreter", kind=KIND_CODE, backend_type="modal-python",
|
||||
summary="Python sandbox. Execute code and return stdout/stderr.",
|
||||
model="modal-python", latency_s=6.0, category=CATEGORY_BASIC,
|
||||
))
|
||||
cat.extend(_openjarvis_basic_tools())
|
||||
|
||||
return cat
|
||||
|
||||
|
||||
def build_tool_specs(tools: List[ExpertTool]) -> List[Dict[str, object]]:
|
||||
"""Turn a tool list into the OpenAI-style tools JSON the policy sees."""
|
||||
return [t.to_spec() for t in tools]
|
||||
|
||||
|
||||
def tools_by_name(tools: List[ExpertTool]) -> Dict[str, ExpertTool]:
|
||||
return {t.name: t for t in tools}
|
||||
|
||||
|
||||
def sample_tool_config(
|
||||
catalog: List[ExpertTool],
|
||||
*,
|
||||
rng: random.Random,
|
||||
min_tools: int = 4,
|
||||
max_tools: Optional[int] = None,
|
||||
price_jitter: float = 0.0,
|
||||
) -> List[ExpertTool]:
|
||||
"""Sample a random tool subset with optional price randomization (§3.3).
|
||||
|
||||
Guarantees at least one ``model`` tool and at least one non-model (basic)
|
||||
tool so every instance can both reason and act. ``price_jitter`` (e.g. 0.5)
|
||||
multiplies each model's prices by a per-tool factor drawn uniformly from
|
||||
``[1-jitter, 1+jitter]``, modeling heterogeneous pricing across users.
|
||||
Deterministic given ``rng``.
|
||||
"""
|
||||
if not catalog:
|
||||
raise ValueError("empty catalog")
|
||||
models = [t for t in catalog if t.kind == KIND_MODEL]
|
||||
basics = [t for t in catalog if t.kind != KIND_MODEL]
|
||||
if not models:
|
||||
raise ValueError("catalog has no model tools")
|
||||
|
||||
hi = max_tools if max_tools is not None else len(catalog)
|
||||
hi = min(hi, len(catalog))
|
||||
lo = min(max(min_tools, 2), hi)
|
||||
k = rng.randint(lo, hi)
|
||||
|
||||
# Always include >=1 model; include >=1 basic if any exist.
|
||||
chosen: List[ExpertTool] = [rng.choice(models)]
|
||||
if basics:
|
||||
chosen.append(rng.choice(basics))
|
||||
pool = [t for t in catalog if t not in chosen]
|
||||
rng.shuffle(pool)
|
||||
for t in pool:
|
||||
if len(chosen) >= k:
|
||||
break
|
||||
chosen.append(t)
|
||||
|
||||
# Re-order to catalog order for stable specs.
|
||||
order = {t.name: i for i, t in enumerate(catalog)}
|
||||
chosen.sort(key=lambda t: order[t.name])
|
||||
|
||||
if price_jitter > 0.0:
|
||||
jittered: List[ExpertTool] = []
|
||||
for t in chosen:
|
||||
if t.kind == KIND_MODEL and (t.price_in or t.price_out):
|
||||
f = rng.uniform(1.0 - price_jitter, 1.0 + price_jitter)
|
||||
jittered.append(ExpertTool(
|
||||
name=t.name, kind=t.kind, backend_type=t.backend_type,
|
||||
summary=t.summary, model=t.model, base_url=t.base_url,
|
||||
price_in=round(t.price_in * f, 4),
|
||||
price_out=round(t.price_out * f, 4),
|
||||
latency_s=t.latency_s,
|
||||
))
|
||||
else:
|
||||
jittered.append(t)
|
||||
return jittered
|
||||
return chosen
|
||||
|
||||
|
||||
def to_worker_dict(tool: ExpertTool) -> Dict[str, object]:
|
||||
"""Convert a tool into the worker dict ``toolorchestra._call_worker`` eats."""
|
||||
d: Dict[str, object] = {
|
||||
"name": tool.name,
|
||||
"type": tool.backend_type,
|
||||
"model": tool.model,
|
||||
}
|
||||
if tool.base_url:
|
||||
d["base_url"] = tool.base_url
|
||||
return d
|
||||
|
||||
|
||||
__all__ = [
|
||||
"CATEGORY_BASIC",
|
||||
"CATEGORY_CLOUD_FRONTIER",
|
||||
"CATEGORY_GENERALIST",
|
||||
"CATEGORY_LOCAL_OSS",
|
||||
"CATEGORY_SPECIALIZED",
|
||||
"ExpertTool",
|
||||
"KIND_CODE",
|
||||
"KIND_LOCAL_SEARCH",
|
||||
"KIND_MODEL",
|
||||
"KIND_TOOL",
|
||||
"KIND_WEB_SEARCH",
|
||||
"build_tool_specs",
|
||||
"default_catalog",
|
||||
"openjarvis_tool",
|
||||
"orchestrator_catalog",
|
||||
"sample_tool_config",
|
||||
"to_worker_dict",
|
||||
"tools_by_name",
|
||||
]
|
||||
File diff suppressed because it is too large
Load Diff
@@ -0,0 +1,20 @@
|
||||
"""ToolOrchestraAgent package (split from the former toolorchestra.py module).
|
||||
|
||||
Importing this package registers the agent and re-exports the public surface,
|
||||
so ``from openjarvis.agents.hybrid.toolorchestra import ToolOrchestraAgent``
|
||||
keeps working unchanged. Submodules:
|
||||
|
||||
prompts — system prompts, RL tool specs / arg schema
|
||||
experts — slot -> backend worker mapping (default + paper-match)
|
||||
sandbox — Tavily search + Modal Python sandbox helpers
|
||||
clients — orchestrator vLLM tool-call client
|
||||
parsing — action / tool-call parsing + user-prompt assembly
|
||||
workers — worker pool resolution + dispatch (_call_worker etc.)
|
||||
agent — ToolOrchestraAgent (the registered agent class)
|
||||
"""
|
||||
|
||||
from __future__ import annotations
|
||||
|
||||
from openjarvis.agents.hybrid.toolorchestra.agent import ToolOrchestraAgent
|
||||
|
||||
__all__ = ["ToolOrchestraAgent"]
|
||||
@@ -0,0 +1,771 @@
|
||||
"""ToolOrchestraAgent — port of NVlabs ToolOrchestra (arXiv:2511.21689).
|
||||
|
||||
Two modes, gated by ``method_cfg.orchestrator_mode``:
|
||||
|
||||
* ``"prompted"`` (default, legacy): a cloud model (Opus etc.) plays the
|
||||
orchestrator, dispatching to a numbered worker pool via JSON
|
||||
``{"action": "call_worker"|"final_answer", ...}`` actions. Useful as
|
||||
a prompted upper-bound reference point — NOT the paper's setup.
|
||||
|
||||
* ``"rl"`` (paper-faithful): the RL-trained ``nvidia/Orchestrator-8B``
|
||||
served on a local vLLM is the orchestrator. It emits OpenAI-style
|
||||
``tool_calls`` (or ``<tool_call>{...}</tool_call>`` text blocks when
|
||||
vLLM's tool parser doesn't catch them) for three expert tools —
|
||||
``enhance_reasoning``, ``answer``, ``search`` — exactly as in the
|
||||
upstream ``evaluation/tools.json``. Each tool's ``model`` arg
|
||||
(``answer-1``, ``reasoner-2``, ``search-3``, …) is mapped to a real
|
||||
backend through ``EXPERT_MODEL_MAPPING`` — by default the frontier
|
||||
Anthropic worker for `*-1` slots, gpt-5-mini for `*-2`, local Qwen
|
||||
for `*-3`. Search routes to the Anthropic server-side web_search.
|
||||
|
||||
We do NOT reproduce the upstream Tavily / FAISS-wiki retriever, the
|
||||
code-interpreter sandbox, or the multi-vLLM mix (Llama-3.3-70B,
|
||||
Qwen-Math, Qwen-Coder); the expert pool collapses onto our existing
|
||||
worker types. Energy-wise, "expert" answers are cloud calls.
|
||||
|
||||
Pipeline per task (RL mode):
|
||||
|
||||
1. Orchestrator-8B reads `Problem: ...\\n\\n{context}\\n\\nChoose an
|
||||
appropriate tool.` with the three tools declared.
|
||||
2. It emits one ``tool_call`` per turn — ``search`` updates the
|
||||
context, ``enhance_reasoning`` appends code/exec output (we run the
|
||||
tool as a plain LLM call, no sandbox — the model just gets prose
|
||||
back), ``answer`` produces the final answer and the loop stops.
|
||||
3. Up to ``max_turns`` (default 8) turns; on parse failure we fall
|
||||
back to the strongest expert worker.
|
||||
|
||||
Prompted-mode pipeline:
|
||||
|
||||
1. Orchestrator (cloud) reads question + numbered worker pool.
|
||||
2. Each turn it emits ``{"action": "call_worker", "worker_id": int,
|
||||
"input": str}`` or ``{"action": "final_answer", "answer": str}``.
|
||||
3. Up to ``max_turns`` (default 6) calls before forcing a final-answer
|
||||
prompt; fallback to strongest worker on parse failure.
|
||||
|
||||
Workers come from ``cfg["workers"]`` or a sensible default pool (local
|
||||
Qwen if vLLM up, plus a web-search tool via Anthropic, Opus 4.7,
|
||||
gpt-5-mini).
|
||||
"""
|
||||
|
||||
|
||||
from __future__ import annotations
|
||||
|
||||
import json
|
||||
import shutil
|
||||
import tempfile
|
||||
from pathlib import Path
|
||||
from typing import Any, Dict, List, Optional, Tuple
|
||||
|
||||
from openjarvis.agents._stubs import AgentContext
|
||||
from openjarvis.agents.hybrid._base import LocalCloudAgent
|
||||
from openjarvis.agents.hybrid._prices import PRICES
|
||||
from openjarvis.agents.hybrid.mini_swe_agent import (
|
||||
_clone_repo,
|
||||
_extract_diff,
|
||||
run_swe_agent_loop,
|
||||
)
|
||||
from openjarvis.core.registry import AgentRegistry
|
||||
|
||||
from openjarvis.agents.hybrid.toolorchestra.prompts import (
|
||||
FORCE_FINAL_PROMPT,
|
||||
ORCHESTRATOR_SYS,
|
||||
RL_ALL_TOOLS,
|
||||
RL_ORCHESTRATOR_SYS,
|
||||
RL_TOOLS_SPEC,
|
||||
)
|
||||
from openjarvis.agents.hybrid.toolorchestra.experts import (
|
||||
_PAPER_CODER_OPENROUTER,
|
||||
_expert_for,
|
||||
_paper_expert_for,
|
||||
)
|
||||
from openjarvis.agents.hybrid.toolorchestra.sandbox import (
|
||||
_call_modal_python,
|
||||
_extract_first_python_block,
|
||||
)
|
||||
from openjarvis.agents.hybrid.toolorchestra.clients import (
|
||||
_call_orchestrator_with_tool_calls,
|
||||
)
|
||||
from openjarvis.agents.hybrid.toolorchestra.parsing import (
|
||||
_build_user_prompt,
|
||||
_extract_final_answer_text,
|
||||
_parse_action,
|
||||
_parse_rl_tool_call,
|
||||
)
|
||||
from openjarvis.agents.hybrid.toolorchestra.workers import (
|
||||
_call_worker,
|
||||
_default_pool,
|
||||
_resolve_worker_pool,
|
||||
_swe_call_worker,
|
||||
)
|
||||
|
||||
@AgentRegistry.register("toolorchestra")
|
||||
class ToolOrchestraAgent(LocalCloudAgent):
|
||||
"""Multi-turn dispatcher over a mixed worker pool.
|
||||
|
||||
Two modes (see module docstring): ``method_cfg.orchestrator_mode``
|
||||
is ``"prompted"`` (default, cloud-as-orchestrator) or ``"rl"``
|
||||
(paper-faithful, drives ``nvidia/Orchestrator-8B`` on a local vLLM).
|
||||
"""
|
||||
|
||||
agent_id = "toolorchestra"
|
||||
|
||||
def __init__(self, *args: Any, **kwargs: Any) -> None:
|
||||
super().__init__(*args, **kwargs)
|
||||
# Validate `method_cfg.worker_pool` early — surfaces config errors
|
||||
# at agent construction rather than on the first task. No-op when
|
||||
# the override is absent.
|
||||
if self._cfg.get("worker_pool") is not None:
|
||||
_resolve_worker_pool(
|
||||
self._cfg,
|
||||
self._local_model,
|
||||
self._local_endpoint,
|
||||
self._cloud_model,
|
||||
self._cloud_endpoint,
|
||||
)
|
||||
# Validate `orchestrator_mode` (typo-checked here, not on first task).
|
||||
mode = str(self._cfg.get("orchestrator_mode", "prompted")).lower()
|
||||
if mode not in ("prompted", "rl"):
|
||||
raise ValueError(
|
||||
f"toolorchestra: orchestrator_mode must be 'prompted' or 'rl'; "
|
||||
f"got {mode!r}"
|
||||
)
|
||||
|
||||
def _run_paradigm(
|
||||
self,
|
||||
input: str,
|
||||
context: Optional[AgentContext],
|
||||
**kwargs: Any,
|
||||
) -> Tuple[str, Dict[str, Any]]:
|
||||
mode = str(self._cfg.get("orchestrator_mode", "prompted")).lower()
|
||||
if mode == "rl":
|
||||
return self._run_rl(input, context, **kwargs)
|
||||
return self._run_prompted(input, context, **kwargs)
|
||||
|
||||
# ------------------------------------------------------------------
|
||||
# Legacy prompted-orchestrator path.
|
||||
# ------------------------------------------------------------------
|
||||
def _run_prompted(
|
||||
self,
|
||||
input: str,
|
||||
context: Optional[AgentContext],
|
||||
**kwargs: Any,
|
||||
) -> Tuple[str, Dict[str, Any]]:
|
||||
cfg = self._cfg
|
||||
question = input
|
||||
# Resolution order (strict replace, no merge):
|
||||
# 1. `cfg["workers"]` — legacy direct override, used by tests.
|
||||
# 2. `cfg["worker_pool"]` — cell-config override; validated +
|
||||
# $local/$cloud substituted.
|
||||
# 3. `_default_pool(...)` — heterogeneous default.
|
||||
if cfg.get("workers"):
|
||||
workers = cfg["workers"]
|
||||
else:
|
||||
workers = _resolve_worker_pool(
|
||||
cfg,
|
||||
self._local_model,
|
||||
self._local_endpoint,
|
||||
self._cloud_model,
|
||||
self._cloud_endpoint,
|
||||
)
|
||||
if not workers:
|
||||
raise RuntimeError("toolorchestra: empty worker pool")
|
||||
|
||||
max_turns = int(cfg.get("max_turns", 6))
|
||||
orch_max_tokens = int(cfg.get("orchestrator_max_tokens", 1024))
|
||||
|
||||
task_meta = (context.metadata.get("task") if context is not None else {}) or {}
|
||||
swe_mode = (
|
||||
bool(cfg.get("swe_use_agent_loop"))
|
||||
and bool(task_meta.get("problem_statement"))
|
||||
and bool(task_meta.get("repo"))
|
||||
and bool(task_meta.get("base_commit"))
|
||||
)
|
||||
shared_workdir: Optional[Path] = None
|
||||
if swe_mode:
|
||||
shared_workdir = Path(tempfile.mkdtemp(
|
||||
prefix=f"toolorch-swe-{task_meta.get('task_id','x')}-"
|
||||
))
|
||||
try:
|
||||
_clone_repo(task_meta["repo"], task_meta["base_commit"], shared_workdir)
|
||||
except Exception:
|
||||
shutil.rmtree(shared_workdir, ignore_errors=True)
|
||||
raise
|
||||
self.record_trace_event({
|
||||
"kind": "toolorchestra_swe_workdir",
|
||||
"workdir": str(shared_workdir),
|
||||
"repo": task_meta["repo"],
|
||||
"base_commit": task_meta["base_commit"],
|
||||
})
|
||||
|
||||
# try/finally guards ``shared_workdir`` against exceptions raised
|
||||
# anywhere in the turn loop, the worker calls, the fallback, or
|
||||
# the diff-extraction step. Without this, at n=500 SWE-bench an
|
||||
# exception leaves hundreds of MB of cloned repos in tempdir.
|
||||
try:
|
||||
history: List[Dict[str, Any]] = []
|
||||
tokens_local = 0
|
||||
tokens_cloud = 0
|
||||
cost = 0.0
|
||||
n_web_searches_total = 0
|
||||
# tool_calls: bash turns from SWE subloops + web_search uses
|
||||
# from GAIA. Orchestrator dispatch turns are NOT counted (they
|
||||
# produce text only — calling a worker is one tool call's worth
|
||||
# of "delegation" but the actual tool action happens inside).
|
||||
tool_calls = 0
|
||||
final_answer: Optional[str] = None
|
||||
forced_final = False
|
||||
parse_failures = 0
|
||||
|
||||
for turn in range(1, max_turns + 1):
|
||||
sys_prompt = ORCHESTRATOR_SYS
|
||||
if turn == max_turns and final_answer is None:
|
||||
sys_prompt = ORCHESTRATOR_SYS + "\n\n" + FORCE_FINAL_PROMPT
|
||||
forced_final = True
|
||||
|
||||
user = _build_user_prompt(question, workers, history)
|
||||
text, o_in, o_out = self._call_cloud(
|
||||
user=user,
|
||||
system=sys_prompt,
|
||||
max_tokens=orch_max_tokens,
|
||||
temperature=0.0,
|
||||
)
|
||||
tokens_cloud += o_in + o_out
|
||||
cost += self.cost_usd(self._cloud_model, o_in, o_out)
|
||||
|
||||
action = _parse_action(text)
|
||||
history.append({
|
||||
"role": "orchestrator", "turn": turn, "raw": text, "action": action,
|
||||
})
|
||||
self.record_trace_event({
|
||||
"kind": "toolorchestra_action",
|
||||
"turn": turn,
|
||||
"action": action,
|
||||
"raw": text,
|
||||
})
|
||||
|
||||
if action is None:
|
||||
parse_failures += 1
|
||||
if parse_failures >= 2 or forced_final:
|
||||
final_answer = _extract_final_answer_text(text)
|
||||
break
|
||||
continue
|
||||
|
||||
kind = action.get("action")
|
||||
if kind == "final_answer":
|
||||
final_answer = str(action.get("answer", "")).strip()
|
||||
break
|
||||
if kind == "call_worker":
|
||||
wid = action.get("worker_id")
|
||||
w_input = action.get("input", "")
|
||||
if not isinstance(wid, int) or not (0 <= wid < len(workers)):
|
||||
parse_failures += 1
|
||||
if parse_failures >= 2 or forced_final:
|
||||
final_answer = _extract_final_answer_text(text)
|
||||
break
|
||||
continue
|
||||
worker = workers[wid]
|
||||
if swe_mode and shared_workdir is not None:
|
||||
(w_text, w_in, w_out, is_local, extra_cost,
|
||||
n_searches, bash_turns) = (
|
||||
_swe_call_worker(
|
||||
worker, str(w_input), cfg, task_meta,
|
||||
shared_workdir, turn,
|
||||
)
|
||||
)
|
||||
tool_calls += bash_turns
|
||||
else:
|
||||
w_text, w_in, w_out, is_local, extra_cost, n_searches = (
|
||||
_call_worker(worker, str(w_input), cfg)
|
||||
)
|
||||
if is_local:
|
||||
tokens_local += w_in + w_out
|
||||
else:
|
||||
tokens_cloud += w_in + w_out
|
||||
cost += self.cost_usd(worker["model"], w_in, w_out) + extra_cost
|
||||
n_web_searches_total += n_searches
|
||||
tool_calls += n_searches
|
||||
history.append({
|
||||
"role": "worker",
|
||||
"turn": turn,
|
||||
"worker_id": wid,
|
||||
"worker_name": worker["name"],
|
||||
"worker_model": worker["model"],
|
||||
"output": w_text,
|
||||
"tokens_in": w_in,
|
||||
"tokens_out": w_out,
|
||||
"n_web_searches": n_searches,
|
||||
})
|
||||
continue
|
||||
# Unknown action kind — treat as parse failure.
|
||||
parse_failures += 1
|
||||
|
||||
if final_answer is None:
|
||||
# Hard fallback: call the strongest non-search worker directly.
|
||||
# "Strongest" = highest output-token price in `_prices.PRICES`,
|
||||
# which tracks model capability tier closely enough for this.
|
||||
# Search workers are excluded — they answer fact-lookup
|
||||
# questions, not synthesis.
|
||||
non_search = [
|
||||
w for w in workers if w.get("type") != "anthropic-web-search"
|
||||
] or workers
|
||||
worker = max(
|
||||
non_search,
|
||||
key=lambda w: PRICES.get(w.get("model", ""), (0.0, 0.0))[1],
|
||||
)
|
||||
if swe_mode and shared_workdir is not None:
|
||||
(ans, w_in, w_out, is_local, extra_cost, _,
|
||||
bash_turns) = _swe_call_worker(
|
||||
worker, question, cfg, task_meta,
|
||||
shared_workdir, max_turns + 1,
|
||||
)
|
||||
tool_calls += bash_turns
|
||||
else:
|
||||
ans, w_in, w_out, is_local, extra_cost, _ = _call_worker(
|
||||
worker, question, cfg
|
||||
)
|
||||
if is_local:
|
||||
tokens_local += w_in + w_out
|
||||
else:
|
||||
tokens_cloud += w_in + w_out
|
||||
cost += self.cost_usd(worker["model"], w_in, w_out) + extra_cost
|
||||
history.append({
|
||||
"role": "worker",
|
||||
"turn": max_turns + 1,
|
||||
"worker_id": worker["id"],
|
||||
"worker_name": worker["name"],
|
||||
"worker_model": worker["model"],
|
||||
"output": ans,
|
||||
"tokens_in": w_in,
|
||||
"tokens_out": w_out,
|
||||
"fallback": True,
|
||||
})
|
||||
final_answer = ans
|
||||
|
||||
# In SWE mode, the authoritative output is the working-tree diff —
|
||||
# frame it (the runner extracts it via the scorer's ```diff fence).
|
||||
if swe_mode and shared_workdir is not None:
|
||||
patch = _extract_diff(shared_workdir)
|
||||
if patch.strip():
|
||||
final_answer = (
|
||||
f"{final_answer}\n\n```diff\n{patch}```"
|
||||
if final_answer else f"```diff\n{patch}```"
|
||||
)
|
||||
|
||||
meta = {
|
||||
"tokens_local": tokens_local,
|
||||
"tokens_cloud": tokens_cloud,
|
||||
"cost_usd": cost,
|
||||
"turns": len([h for h in history if h["role"] == "orchestrator"]),
|
||||
"web_search_uses": n_web_searches_total,
|
||||
"tool_calls": int(tool_calls),
|
||||
"traces": {
|
||||
"history": history,
|
||||
"forced_final": forced_final,
|
||||
"parse_failures": parse_failures,
|
||||
"workers": workers,
|
||||
"n_web_searches": n_web_searches_total,
|
||||
"note": (
|
||||
"inference-only port; the RL-trained Nemotron-Orchestrator-8B "
|
||||
"is NOT in the loop. Results are preliminary."
|
||||
),
|
||||
},
|
||||
}
|
||||
return final_answer, meta
|
||||
finally:
|
||||
if shared_workdir is not None:
|
||||
shutil.rmtree(shared_workdir, ignore_errors=True)
|
||||
|
||||
# ------------------------------------------------------------------
|
||||
# Paper-faithful Orchestrator-8B path.
|
||||
# ------------------------------------------------------------------
|
||||
def _run_rl(
|
||||
self,
|
||||
input: str,
|
||||
context: Optional[AgentContext],
|
||||
**kwargs: Any,
|
||||
) -> Tuple[str, Dict[str, Any]]:
|
||||
cfg = self._cfg
|
||||
question = input
|
||||
|
||||
# Orchestrator endpoint / model (where the RL'd 8B lives).
|
||||
orch_endpoint = str(
|
||||
cfg.get("orchestrator_endpoint", "http://localhost:8003/v1")
|
||||
)
|
||||
orch_model = str(cfg.get("orchestrator_model", "orchestrator-8b"))
|
||||
max_turns = int(cfg.get("max_turns", 8))
|
||||
orch_max_tokens = int(cfg.get("orchestrator_max_tokens", 4096))
|
||||
orch_temp = float(cfg.get("orchestrator_temperature", 1.0))
|
||||
|
||||
# Paper-match pool toggle (2026-05-19). When set, `_paper_expert_for`
|
||||
# replaces `_expert_for` and `enhance_reasoning` is post-processed
|
||||
# through a Modal Python sandbox. See module docstring + paper-match
|
||||
# doc at `docs/26.5.19/toolorchestra-papermatch.md`.
|
||||
paper_mode = str(cfg.get("pool", "")).lower() == "paper"
|
||||
|
||||
# SWE-bench detection: same gate as the prompted path. Requires
|
||||
# `method_cfg.swe_use_agent_loop = true` AND the task carries the
|
||||
# SWE-bench fields. When active, the `enhance_reasoning` and
|
||||
# `answer` workers route through `run_swe_agent_loop` on a shared
|
||||
# workdir; search workers stay one-shot. At end, the working-tree
|
||||
# diff is appended to final_answer so `_score_swebench` can extract
|
||||
# it via the ```diff fence.
|
||||
task_meta = (context.metadata.get("task") if context is not None else {}) or {}
|
||||
swe_mode = (
|
||||
bool(cfg.get("swe_use_agent_loop"))
|
||||
and bool(task_meta.get("problem_statement"))
|
||||
and bool(task_meta.get("repo"))
|
||||
and bool(task_meta.get("base_commit"))
|
||||
)
|
||||
shared_workdir: Optional[Path] = None
|
||||
if swe_mode:
|
||||
shared_workdir = Path(tempfile.mkdtemp(
|
||||
prefix=f"toolorch-rl-swe-{task_meta.get('task_id','x')}-"
|
||||
))
|
||||
try:
|
||||
_clone_repo(task_meta["repo"], task_meta["base_commit"], shared_workdir)
|
||||
except Exception:
|
||||
shutil.rmtree(shared_workdir, ignore_errors=True)
|
||||
raise
|
||||
self.record_trace_event({
|
||||
"kind": "toolorchestra_rl_swe_workdir",
|
||||
"workdir": str(shared_workdir),
|
||||
"repo": task_meta["repo"],
|
||||
"base_commit": task_meta["base_commit"],
|
||||
})
|
||||
|
||||
# ``context_str`` mirrors the upstream's running context — accumulates
|
||||
# search documents and code/exec snippets across turns. We keep this
|
||||
# as a single string for prompt simplicity; the upstream uses
|
||||
# tokenized cutoffs (we cap at ~24k chars instead).
|
||||
context_str = ""
|
||||
doc_list: List[str] = []
|
||||
history: List[Dict[str, Any]] = []
|
||||
tokens_local = 0
|
||||
tokens_cloud = 0
|
||||
cost = 0.0
|
||||
n_web_searches_total = 0
|
||||
tool_calls = 0
|
||||
final_answer: Optional[str] = None
|
||||
parse_failures = 0
|
||||
|
||||
# Single outer try/finally guards `shared_workdir` against any
|
||||
# exception in the orchestrator loop, the post-loop fallback, or
|
||||
# the diff-extraction step. Matches the prompted path's pattern.
|
||||
try:
|
||||
for turn in range(1, max_turns + 1):
|
||||
user = (
|
||||
f"Problem: {question}\n\n{context_str}\n\n"
|
||||
"Choose an appropriate tool."
|
||||
)
|
||||
|
||||
# Orchestrator-8B served on local vLLM. We pass the three NVlabs
|
||||
# tools verbatim. In paper-mode we use the local helper so we
|
||||
# get the SDK-level ``tool_calls`` object back — `_call_vllm`
|
||||
# returns just text and loses the call when vLLM's parser
|
||||
# caught it. Orchestrator-8B emits its routing decision in the
|
||||
# OpenAI-native ``tool_calls`` array with an empty text body,
|
||||
# so the legacy `_call_vllm` path saw nothing and silently fell
|
||||
# through to the answer-1 fallback (parse_failures: 2 on every
|
||||
# non-opus-gaia cell — see docs/reports/toolorchestra.md). Both
|
||||
# modes now use `_call_orchestrator_with_tool_calls` so the
|
||||
# parser can read structured tool calls; the text-tag path in
|
||||
# `_parse_rl_tool_call` is still the fallback when `tool_calls`
|
||||
# is empty.
|
||||
text, o_in, o_out, sdk_tool_calls = _call_orchestrator_with_tool_calls(
|
||||
orch_model,
|
||||
orch_endpoint,
|
||||
user=user,
|
||||
system=RL_ORCHESTRATOR_SYS,
|
||||
max_tokens=orch_max_tokens,
|
||||
temperature=orch_temp,
|
||||
tools=RL_TOOLS_SPEC,
|
||||
)
|
||||
self.record_trace_event({
|
||||
"kind": "vllm",
|
||||
"role": "orchestrator",
|
||||
"model": orch_model,
|
||||
"endpoint": orch_endpoint,
|
||||
"system": RL_ORCHESTRATOR_SYS,
|
||||
"user": user,
|
||||
"response": text,
|
||||
"tool_calls": [
|
||||
{
|
||||
"id": getattr(tc, "id", None),
|
||||
"type": getattr(tc, "type", None),
|
||||
"function": {
|
||||
"name": getattr(getattr(tc, "function", None), "name", None),
|
||||
"arguments": getattr(getattr(tc, "function", None), "arguments", None),
|
||||
},
|
||||
}
|
||||
for tc in (sdk_tool_calls or [])
|
||||
],
|
||||
"tokens_in": o_in,
|
||||
"tokens_out": o_out,
|
||||
})
|
||||
tokens_local += o_in + o_out
|
||||
|
||||
action = _parse_rl_tool_call(text, sdk_tool_calls)
|
||||
history.append({
|
||||
"role": "orchestrator", "turn": turn, "raw": text, "action": action,
|
||||
})
|
||||
self.record_trace_event({
|
||||
"kind": "toolorchestra_rl_action",
|
||||
"turn": turn,
|
||||
"action": action,
|
||||
"raw": text,
|
||||
})
|
||||
|
||||
if action is None:
|
||||
parse_failures += 1
|
||||
if parse_failures >= 2:
|
||||
break
|
||||
continue
|
||||
|
||||
name = action["name"]
|
||||
args = action.get("arguments", {})
|
||||
slot = args.get("model", "")
|
||||
|
||||
# Validate against the upstream tool/arg schema.
|
||||
valid = name in RL_ALL_TOOLS and isinstance(slot, str) and (
|
||||
slot in RL_ALL_TOOLS[name]["model"]
|
||||
)
|
||||
if not valid:
|
||||
parse_failures += 1
|
||||
if parse_failures >= 2:
|
||||
break
|
||||
# Replay with a softer nudge in the context.
|
||||
context_str += (
|
||||
f"\n[Orchestrator emitted invalid tool call "
|
||||
f"name={name!r} slot={slot!r} — try again.]\n"
|
||||
)
|
||||
continue
|
||||
|
||||
# Paper-match (`method_cfg.pool == "paper"`) routes through
|
||||
# the Tavily/OpenRouter/Modal pool instead of the default
|
||||
# Anthropic-web-search-driven mapping. For `search` this
|
||||
# also forces the worker prompt to a raw query string
|
||||
# (Tavily takes a single search string, not a chat-style
|
||||
# framing).
|
||||
if paper_mode:
|
||||
worker = _paper_expert_for(
|
||||
slot, self._local_model, self._local_endpoint,
|
||||
self._cloud_model, self._cloud_endpoint,
|
||||
)
|
||||
# In paper mode, `enhance_reasoning` is always the coder
|
||||
# specialist regardless of the orchestrator's chosen tier.
|
||||
# The coder is then expected to emit a python block which
|
||||
# we exec in Modal (below).
|
||||
if name == "enhance_reasoning":
|
||||
worker = {
|
||||
"name": f"coder:{slot}",
|
||||
"type": "openrouter",
|
||||
"model": _PAPER_CODER_OPENROUTER,
|
||||
}
|
||||
else:
|
||||
worker = _expert_for(
|
||||
slot, self._local_model, self._local_endpoint, self._cloud_model,
|
||||
self._cloud_endpoint,
|
||||
)
|
||||
|
||||
# Dispatch — the orchestrator only conveys a tool/model
|
||||
# choice, NOT a question rewrite; the prompt we send the
|
||||
# expert is the same context the orchestrator saw, framed
|
||||
# appropriately for the tool.
|
||||
if name == "search":
|
||||
if paper_mode:
|
||||
# Tavily takes a query string. Orchestrator-8B often
|
||||
# emits an extra `query` arg (not in the upstream
|
||||
# schema but useful) — prefer it; else fall back to
|
||||
# the raw question.
|
||||
q = args.get("query")
|
||||
w_input = q if isinstance(q, str) and q.strip() else question
|
||||
else:
|
||||
w_input = (
|
||||
f"Search the web to gather information that helps answer:\n"
|
||||
f"{question}\n\nCurrent context:\n{context_str or '(empty)'}"
|
||||
)
|
||||
elif name == "enhance_reasoning":
|
||||
if paper_mode:
|
||||
w_input = (
|
||||
f"Problem: {question}\n\nContext:\n{context_str or '(empty)'}\n\n"
|
||||
"Write a short Python script that computes intermediate "
|
||||
"results which help answer the problem. Output ONLY the "
|
||||
"code inside one ```python ... ``` fenced block. Print "
|
||||
"any results you derive using `print(...)`. The script "
|
||||
"must run with the Python stdlib only — no extra pip "
|
||||
"installs."
|
||||
)
|
||||
else:
|
||||
w_input = (
|
||||
f"Problem: {question}\n\nContext:\n{context_str or '(empty)'}\n\n"
|
||||
"Reason carefully. Outline the key intermediate steps and any "
|
||||
"computations or facts you can derive. Do NOT give a final "
|
||||
"answer — the orchestrator will collect your reasoning and "
|
||||
"call the answer tool next."
|
||||
)
|
||||
else: # name == "answer"
|
||||
w_input = (
|
||||
f"Problem: {question}\n\nContext:\n{context_str or '(empty)'}\n\n"
|
||||
"Provide the final answer to the user. Respect any "
|
||||
"answer-format rules in the question (e.g. GAIA's "
|
||||
"FINAL ANSWER: <value> convention)."
|
||||
)
|
||||
|
||||
# SWE mode: route enhance_reasoning / answer workers through
|
||||
# the SWE agent loop on the shared workdir so they can read
|
||||
# files, run tests, and edit the working tree. Search workers
|
||||
# stay one-shot (no agent loop). The `_swe_call_worker`
|
||||
# one-shot fallbacks (openai-typed workers, search) return
|
||||
# bash_turns=0; vllm/anthropic-typed workers run the loop.
|
||||
bash_turns = 0
|
||||
if swe_mode and shared_workdir is not None and name != "search":
|
||||
(w_text, w_in, w_out, is_local, extra_cost,
|
||||
n_searches, bash_turns) = _swe_call_worker(
|
||||
worker, w_input, cfg, task_meta, shared_workdir, turn,
|
||||
)
|
||||
else:
|
||||
w_text, w_in, w_out, is_local, extra_cost, n_searches = _call_worker(
|
||||
worker, w_input, cfg
|
||||
)
|
||||
if is_local:
|
||||
tokens_local += w_in + w_out
|
||||
else:
|
||||
tokens_cloud += w_in + w_out
|
||||
cost += self.cost_usd(worker["model"], w_in, w_out) + extra_cost
|
||||
n_web_searches_total += n_searches
|
||||
# SWE bash turns count as tool calls (each one is a $BASH block
|
||||
# the agent executed). On non-SWE turns fall back to the
|
||||
# original "at least one expert call" accounting.
|
||||
tool_calls += bash_turns if bash_turns > 0 else max(1, n_searches)
|
||||
|
||||
# Paper-match: pipe coder output through a Modal sandbox so
|
||||
# `enhance_reasoning` actually executes the code the coder
|
||||
# wrote. Append the exec output to the worker's text. No-op
|
||||
# when no python block is found.
|
||||
modal_exec_output: Optional[str] = None
|
||||
modal_exec_rc: Optional[int] = None
|
||||
if (paper_mode and name == "enhance_reasoning"
|
||||
and not swe_mode):
|
||||
code = _extract_first_python_block(w_text)
|
||||
if code:
|
||||
timeout_s = int(cfg.get("modal_python_timeout_s", 60))
|
||||
modal_exec_output, modal_exec_rc = _call_modal_python(
|
||||
code, timeout_s=timeout_s,
|
||||
)
|
||||
tool_calls += 1
|
||||
w_text = (
|
||||
f"{w_text}\n\n[modal-python stdout/stderr "
|
||||
f"(rc={modal_exec_rc})]\n{modal_exec_output}"
|
||||
)
|
||||
|
||||
history.append({
|
||||
"role": "worker",
|
||||
"turn": turn,
|
||||
"tool": name,
|
||||
"slot": slot,
|
||||
"worker_model": worker["model"],
|
||||
"worker_type": worker["type"],
|
||||
"output": w_text,
|
||||
"tokens_in": w_in,
|
||||
"tokens_out": w_out,
|
||||
"n_web_searches": n_searches,
|
||||
"bash_turns": bash_turns,
|
||||
"modal_exec_rc": modal_exec_rc,
|
||||
})
|
||||
|
||||
# Update accumulated context for the next turn.
|
||||
if name == "search":
|
||||
# Treat the search worker's response as a document.
|
||||
doc_list.append(w_text)
|
||||
ctx_docs = "\n\n".join(
|
||||
f"Doc {i+1}: {d}" for i, d in enumerate(doc_list)
|
||||
)
|
||||
# Crude char-level cap mirrors the upstream's ~24k token cap.
|
||||
context_str = ("Documents:\n" + ctx_docs)[-24000:]
|
||||
elif name == "enhance_reasoning":
|
||||
snippet = f"\n\nReasoning/exec output:\n{w_text}"
|
||||
context_str = (context_str + snippet)[-24000:]
|
||||
else: # answer
|
||||
final_answer = w_text.strip()
|
||||
break
|
||||
|
||||
if final_answer is None:
|
||||
# Hard fallback: ask the frontier worker directly. In SWE
|
||||
# mode route this final call through the agent loop too so
|
||||
# it can still touch the workdir and emit a diff.
|
||||
expert_fn = _paper_expert_for if paper_mode else _expert_for
|
||||
worker = expert_fn(
|
||||
"answer-1", self._local_model, self._local_endpoint,
|
||||
self._cloud_model, self._cloud_endpoint,
|
||||
)
|
||||
fb_bash_turns = 0
|
||||
if swe_mode and shared_workdir is not None:
|
||||
(ans, w_in, w_out, is_local, extra_cost,
|
||||
_, fb_bash_turns) = _swe_call_worker(
|
||||
worker, question, cfg, task_meta,
|
||||
shared_workdir, max_turns + 1,
|
||||
)
|
||||
tool_calls += fb_bash_turns
|
||||
else:
|
||||
ans, w_in, w_out, is_local, extra_cost, _ = _call_worker(
|
||||
worker, question, cfg
|
||||
)
|
||||
if is_local:
|
||||
tokens_local += w_in + w_out
|
||||
else:
|
||||
tokens_cloud += w_in + w_out
|
||||
cost += self.cost_usd(worker["model"], w_in, w_out) + extra_cost
|
||||
history.append({
|
||||
"role": "worker",
|
||||
"turn": max_turns + 1,
|
||||
"tool": "answer",
|
||||
"slot": "answer-1",
|
||||
"worker_model": worker["model"],
|
||||
"worker_type": worker["type"],
|
||||
"output": ans,
|
||||
"tokens_in": w_in,
|
||||
"tokens_out": w_out,
|
||||
"bash_turns": fb_bash_turns,
|
||||
"fallback": True,
|
||||
})
|
||||
final_answer = ans
|
||||
|
||||
# In SWE mode, the authoritative output is the working-tree diff —
|
||||
# frame it so `_score_swebench`'s extract_patch picks it up.
|
||||
if swe_mode and shared_workdir is not None:
|
||||
patch = _extract_diff(shared_workdir)
|
||||
if patch.strip():
|
||||
final_answer = (
|
||||
f"{final_answer}\n\n```diff\n{patch}```"
|
||||
if final_answer else f"```diff\n{patch}```"
|
||||
)
|
||||
|
||||
meta = {
|
||||
"tokens_local": tokens_local,
|
||||
"tokens_cloud": tokens_cloud,
|
||||
"cost_usd": cost,
|
||||
"turns": len([h for h in history if h["role"] == "orchestrator"]),
|
||||
"web_search_uses": n_web_searches_total,
|
||||
"tool_calls": int(tool_calls),
|
||||
"traces": {
|
||||
"history": history,
|
||||
"parse_failures": parse_failures,
|
||||
"orchestrator_model": orch_model,
|
||||
"orchestrator_endpoint": orch_endpoint,
|
||||
"mode": "rl",
|
||||
"pool": "paper" if paper_mode else "default",
|
||||
"swe_mode": swe_mode,
|
||||
"note": (
|
||||
"RL-trained nvidia/Orchestrator-8B as orchestrator. "
|
||||
"Expert pool collapses Tavily/FAISS/Qwen-Math/Coder onto "
|
||||
"our hybrid worker types — see toolorchestra.py docstring."
|
||||
),
|
||||
},
|
||||
}
|
||||
return final_answer, meta
|
||||
finally:
|
||||
if shared_workdir is not None:
|
||||
shutil.rmtree(shared_workdir, ignore_errors=True)
|
||||
|
||||
|
||||
__all__ = ["ToolOrchestraAgent"]
|
||||
@@ -0,0 +1,50 @@
|
||||
"""Orchestrator vLLM tool-call client for ToolOrchestraAgent."""
|
||||
|
||||
from __future__ import annotations
|
||||
|
||||
from typing import Any, Dict, List, Tuple
|
||||
|
||||
def _call_orchestrator_with_tool_calls(
|
||||
model: str,
|
||||
endpoint: str,
|
||||
*,
|
||||
user: str,
|
||||
system: str,
|
||||
max_tokens: int,
|
||||
temperature: float,
|
||||
tools: List[Dict[str, Any]],
|
||||
timeout: float = 600.0,
|
||||
) -> Tuple[str, int, int, Any]:
|
||||
"""Orchestrator-aware vLLM call. Returns (text, p_tok, c_tok, tool_calls).
|
||||
|
||||
Mirrors ``LocalCloudAgent._call_vllm`` but ALSO surfaces the SDK-level
|
||||
``tool_calls`` object so the RL-mode parser can match against it
|
||||
directly. Otherwise vLLM's tool parser silently swallows the tool call
|
||||
into the SDK field while leaving ``content == ''`` — and the text-tag
|
||||
parser sees nothing, falling through to the answer-1 fallback. (Bug
|
||||
observed 2026-05-19 on the paper-match smoke; same path was buggy on
|
||||
the default pool too, just less reproducibly.)
|
||||
"""
|
||||
from openai import OpenAI
|
||||
|
||||
client = OpenAI(base_url=endpoint, api_key="EMPTY", timeout=timeout)
|
||||
messages = [
|
||||
{"role": "system", "content": system},
|
||||
{"role": "user", "content": user},
|
||||
]
|
||||
resp = client.chat.completions.create(
|
||||
model=model,
|
||||
messages=messages,
|
||||
temperature=temperature,
|
||||
max_tokens=max_tokens,
|
||||
tools=tools,
|
||||
extra_body={"chat_template_kwargs": {"enable_thinking": False}},
|
||||
)
|
||||
choice = resp.choices[0]
|
||||
message = choice.message
|
||||
text = message.content or ""
|
||||
tool_calls = getattr(message, "tool_calls", None)
|
||||
u = resp.usage
|
||||
p = getattr(u, "prompt_tokens", 0) if u else 0
|
||||
c = getattr(u, "completion_tokens", 0) if u else 0
|
||||
return text, p, c, tool_calls
|
||||
@@ -0,0 +1,170 @@
|
||||
"""Slot -> worker expert mapping for ToolOrchestraAgent."""
|
||||
|
||||
from __future__ import annotations
|
||||
|
||||
from typing import Any, Dict, Optional
|
||||
|
||||
# Default model used when an `anthropic-web-search` entry omits `model`.
|
||||
_DEFAULT_WEB_SEARCH_MODEL = "claude-haiku-4-5"
|
||||
|
||||
# Map the orchestrator's `model` slot to a concrete OpenJarvis worker spec.
|
||||
# Tiers ranked by the upstream tools.json table (`*-1` = frontier,
|
||||
# `*-2` = mid, `*-3` = local). math-1 / math-2 collapse onto the same
|
||||
# tiers since we don't have Qwen-Math served.
|
||||
#
|
||||
# Each entry is a callable `(local_model, local_endpoint, cloud_model) -> worker_dict`
|
||||
# so the substitution is deferred until we know the cell's resolved local/cloud
|
||||
# pair. Worker dicts share the schema validated by `_resolve_worker_pool`.
|
||||
|
||||
def _expert_for(slot: str, local_model: Optional[str],
|
||||
local_endpoint: Optional[str],
|
||||
cloud_model: str,
|
||||
cloud_endpoint: str = "anthropic") -> Dict[str, Any]:
|
||||
"""Map an upstream model slot (`answer-1`, `search-3`, …) to a worker spec.
|
||||
|
||||
Routing policy:
|
||||
- `*-1` (frontier tier) -> cloud (`cloud_model`), wtype keyed off
|
||||
`cloud_endpoint` ("anthropic"/"openai"/"gemini")
|
||||
- `*-2` (mid tier) -> cloud `gpt-5-mini` (matches the paper's
|
||||
cost tier for mid OpenAI calls)
|
||||
- `*-3` (local tier) -> local vLLM (`local_model`)
|
||||
- `answer-math-*` -> same tiers as the numeric suffix
|
||||
- `search-*` -> always the Anthropic web_search tool (the
|
||||
upstream uses Tavily; we have web_search)
|
||||
"""
|
||||
if slot.startswith("search"):
|
||||
return {
|
||||
"name": f"search:{slot}",
|
||||
"type": "anthropic-web-search",
|
||||
"model": _DEFAULT_WEB_SEARCH_MODEL,
|
||||
}
|
||||
if slot.endswith("-1") or slot.endswith("-math-1"):
|
||||
ep = (cloud_endpoint or "anthropic").lower()
|
||||
if ep not in ("anthropic", "openai", "gemini"):
|
||||
ep = "anthropic"
|
||||
return {
|
||||
"name": f"frontier:{slot}",
|
||||
"type": ep,
|
||||
"model": cloud_model,
|
||||
}
|
||||
if slot.endswith("-2") or slot.endswith("-math-2"):
|
||||
return {
|
||||
"name": f"mid:{slot}",
|
||||
"type": "openai",
|
||||
"model": "gpt-5-mini",
|
||||
}
|
||||
# `*-3` / `*-4` collapse to local vLLM (paper uses Qwen3-32B etc.;
|
||||
# we substitute whatever local model the cell wired up).
|
||||
if local_model and local_endpoint:
|
||||
return {
|
||||
"name": f"local:{slot}",
|
||||
"type": "vllm",
|
||||
"model": local_model,
|
||||
"base_url": local_endpoint,
|
||||
}
|
||||
# Fallback if no local — gpt-5-mini.
|
||||
return {
|
||||
"name": f"mid-fallback:{slot}",
|
||||
"type": "openai",
|
||||
"model": "gpt-5-mini",
|
||||
}
|
||||
|
||||
|
||||
# ============================================================================
|
||||
# Paper-match expert mapping (2026-05-19).
|
||||
# ============================================================================
|
||||
# Maps the orchestrator's `model` slot to a paper-match worker spec. Differs
|
||||
# from `_expert_for` in that it pulls in OpenRouter-hosted code/math/generalist
|
||||
# models and routes `search` through Tavily, while `enhance_reasoning` is
|
||||
# expected to produce code that the caller pipes through a Modal sandbox
|
||||
# (handled at dispatch time, not here).
|
||||
#
|
||||
# Slot map (paper-faithful where we can; substitutions noted in toolorchestra
|
||||
# paper-match docs `docs/26.5.19/toolorchestra-papermatch.md`):
|
||||
#
|
||||
# reasoner-1 -> GPT-5 (frontier reasoner)
|
||||
# reasoner-2 -> GPT-5-mini (mid)
|
||||
# reasoner-3 -> local Qwen (Orchestrator-8B endpoint also serves this)
|
||||
# answer-1 -> GPT-5
|
||||
# answer-2 -> GPT-5-mini
|
||||
# answer-3 -> Llama-3.3-70B (OpenRouter, generalist tier-3 per spec)
|
||||
# answer-4 -> local Qwen
|
||||
# answer-math-1 -> Qwen-2.5-Coder-32B via OpenRouter
|
||||
# (paper uses Qwen-2.5-Math-72B; not on OpenRouter — see doc)
|
||||
# answer-math-2 -> Qwen-2.5-Coder-32B via OpenRouter
|
||||
# (paper uses Qwen-2.5-Math-7B; not on OpenRouter — see doc)
|
||||
# search-* -> Tavily search (paper)
|
||||
#
|
||||
# `enhance_reasoning` is dispatched through the coder specialist regardless of
|
||||
# slot tier — the orchestrator emits one of `reasoner-{1,2,3}` and the caller
|
||||
# routes the same way in all three cases, then optionally extracts a python
|
||||
# code block and execs it in Modal. (We keep the slot-aware routing inside the
|
||||
# `reasoner-*` map above for parity, but the `enhance_reasoning` tool itself
|
||||
# pins the coder regardless. See `_run_rl_paper` dispatch.)
|
||||
|
||||
_PAPER_CODER_OPENROUTER = "qwen/qwen-2.5-coder-32b-instruct"
|
||||
_PAPER_GENERALIST_TIER3_OPENROUTER = "meta-llama/llama-3.3-70b-instruct"
|
||||
|
||||
|
||||
def _paper_expert_for(
|
||||
slot: str,
|
||||
local_model: Optional[str],
|
||||
local_endpoint: Optional[str],
|
||||
cloud_model: str,
|
||||
cloud_endpoint: str = "openai",
|
||||
) -> Dict[str, Any]:
|
||||
"""Paper-match counterpart of ``_expert_for``.
|
||||
|
||||
Differs from ``_expert_for``:
|
||||
- Search slots go to ``tavily-search`` (not Anthropic web_search).
|
||||
- Tier-3 generalist answer (``answer-3``) routes to Llama-3.3-70B via
|
||||
OpenRouter rather than collapsing onto the local vLLM.
|
||||
- Math slots route to the OpenRouter code specialist (Qwen-2.5-Coder-32B)
|
||||
as a substitute for the unavailable Qwen-2.5-Math-{72B,7B}.
|
||||
- ``reasoner-1`` / ``answer-1`` route to GPT-5 by default (paper).
|
||||
"""
|
||||
if slot.startswith("search"):
|
||||
return {
|
||||
"name": f"tavily:{slot}",
|
||||
"type": "tavily-search",
|
||||
"model": "tavily",
|
||||
}
|
||||
if slot in ("answer-math-1", "answer-math-2"):
|
||||
return {
|
||||
"name": f"math-coder:{slot}",
|
||||
"type": "openrouter",
|
||||
"model": _PAPER_CODER_OPENROUTER,
|
||||
}
|
||||
if slot == "answer-3":
|
||||
return {
|
||||
"name": f"generalist-llama:{slot}",
|
||||
"type": "openrouter",
|
||||
"model": _PAPER_GENERALIST_TIER3_OPENROUTER,
|
||||
}
|
||||
if slot.endswith("-1"):
|
||||
# Tier-1 frontier reasoner / answer — paper uses GPT-5.
|
||||
return {
|
||||
"name": f"frontier:{slot}",
|
||||
"type": "openai",
|
||||
"model": "gpt-5",
|
||||
}
|
||||
if slot.endswith("-2"):
|
||||
return {
|
||||
"name": f"mid:{slot}",
|
||||
"type": "openai",
|
||||
"model": "gpt-5-mini",
|
||||
}
|
||||
# `*-3` / `*-4` collapse onto the local vLLM (the orchestrator endpoint
|
||||
# also serves the local Qwen for the rare local-tier slot).
|
||||
if local_model and local_endpoint:
|
||||
return {
|
||||
"name": f"local:{slot}",
|
||||
"type": "vllm",
|
||||
"model": local_model,
|
||||
"base_url": local_endpoint,
|
||||
}
|
||||
return {
|
||||
"name": f"mid-fallback:{slot}",
|
||||
"type": "openai",
|
||||
"model": "gpt-5-mini",
|
||||
}
|
||||
@@ -0,0 +1,163 @@
|
||||
"""Action / tool-call parsing + prompt assembly for ToolOrchestraAgent."""
|
||||
|
||||
from __future__ import annotations
|
||||
|
||||
import json
|
||||
import re
|
||||
from typing import Any, Dict, List, Optional
|
||||
|
||||
# Regex for ``<tool_call>{...}</tool_call>`` blocks emitted by Orchestrator-8B
|
||||
# when the vLLM tool parser doesn't catch them (e.g. `qwen3_xml` parser on a
|
||||
# hermes-style template). Captures the JSON payload.
|
||||
_TOOL_CALL_TAG_RE = re.compile(
|
||||
r"<tool_call>\s*(\{.*?\})\s*</tool_call>", re.DOTALL
|
||||
)
|
||||
|
||||
# Fallback for a known failure mode: the SFT'd orchestrator often emits its
|
||||
# delegation as ``\boxed{expert_ab12: <sub-question>}`` (math-data habit bleeding
|
||||
# into routing) instead of a real ``<tool_call>`` JSON block. The name before the
|
||||
# colon is a valid (anonymized) tool label; the rest is the sub-question/args.
|
||||
#
|
||||
# Real outputs use several shapes, all handled here:
|
||||
# \boxed{web_search: <query>} -> name, args
|
||||
# \boxed{web_search query: <query>} -> the word "query" is dropped
|
||||
# \boxed{file_read, path: <path>} -> the ", <key>" hint sets the arg key
|
||||
# We require a ``name <optional , key | query> : <args>`` structure, so a genuine
|
||||
# answer like ``\boxed{10}``, ``\boxed{b,e}`` or ``\boxed{The answer is: 42}``
|
||||
# (first token isn't followed by a bare ``:`` / ``, key:`` / ``query:``) is left
|
||||
# alone. ``group("key")`` is the explicit arg key when the model wrote ``, key:``.
|
||||
_BOXED_DELEGATION_RE = re.compile(
|
||||
r"\\boxed\{\s*([A-Za-z_][\w-]*)\s*"
|
||||
r"(?:,\s*(?P<key>\w+)\s*)?" # optional ", key" hint (e.g. file_read, path:)
|
||||
r"(?:query\s*)?" # optional literal "query" word before the colon
|
||||
r":\s*(.+?)\s*\}\s*$",
|
||||
re.DOTALL,
|
||||
)
|
||||
|
||||
|
||||
def _parse_rl_tool_call(content: str, sdk_tool_calls: Any) -> Optional[Dict[str, Any]]:
|
||||
"""Return ``{"name": str, "arguments": dict}`` or None.
|
||||
|
||||
Prefers the SDK-level ``tool_calls`` (when vLLM's parser matched), falls
|
||||
back to scraping ``<tool_call>{...}</tool_call>`` tags from the raw
|
||||
content. We take the first tool call only — Orchestrator-8B was trained
|
||||
to emit exactly one per turn.
|
||||
"""
|
||||
# SDK-level path.
|
||||
if sdk_tool_calls:
|
||||
first = sdk_tool_calls[0]
|
||||
name = getattr(getattr(first, "function", None), "name", None)
|
||||
args_raw = getattr(getattr(first, "function", None), "arguments", None) or "{}"
|
||||
try:
|
||||
args = json.loads(args_raw)
|
||||
except json.JSONDecodeError:
|
||||
args = {}
|
||||
if isinstance(name, str) and isinstance(args, dict):
|
||||
return {"name": name, "arguments": args}
|
||||
# Text-tag fallback.
|
||||
if not isinstance(content, str):
|
||||
return None
|
||||
m = _TOOL_CALL_TAG_RE.search(content)
|
||||
if m:
|
||||
try:
|
||||
obj = json.loads(m.group(1))
|
||||
except json.JSONDecodeError:
|
||||
obj = None
|
||||
if isinstance(obj, dict):
|
||||
name = obj.get("name")
|
||||
args = obj.get("arguments", {})
|
||||
if isinstance(name, str) and isinstance(args, dict):
|
||||
return {"name": name, "arguments": args}
|
||||
# \boxed{name[, key][ query]: <args>} fallback (see _BOXED_DELEGATION_RE).
|
||||
bm = _BOXED_DELEGATION_RE.search(content.strip())
|
||||
if bm:
|
||||
arg_val = bm.groups()[-1].strip()
|
||||
key = bm.group("key") or "input"
|
||||
return {"name": bm.group(1), "arguments": {key: arg_val}}
|
||||
return None
|
||||
|
||||
|
||||
def _build_pool_block(workers: List[Dict[str, Any]]) -> str:
|
||||
return "\n".join(
|
||||
f"Worker {w['id']} ({w['name']}): {w['description']}" for w in workers
|
||||
)
|
||||
|
||||
|
||||
def _build_user_prompt(
|
||||
question: str,
|
||||
workers: List[Dict[str, Any]],
|
||||
history: List[Dict[str, Any]],
|
||||
) -> str:
|
||||
pieces = [
|
||||
f"Worker pool:\n{_build_pool_block(workers)}",
|
||||
f"User question:\n{question}",
|
||||
]
|
||||
if history:
|
||||
pieces.append("Conversation so far (orchestrator turns and worker outputs):")
|
||||
for h in history:
|
||||
if h["role"] == "orchestrator":
|
||||
pieces.append(f"[Orchestrator turn {h['turn']}]\n{h['raw']}")
|
||||
else:
|
||||
pieces.append(
|
||||
f"[Worker {h['worker_id']} ({h['worker_name']}) turn {h['turn']}]\n"
|
||||
f"{h['output']}"
|
||||
)
|
||||
pieces.append(
|
||||
"Emit the next JSON action object now — exactly one object, no prose."
|
||||
)
|
||||
return "\n\n".join(pieces)
|
||||
|
||||
|
||||
def _strip_fences(s: str) -> str:
|
||||
s = s.strip()
|
||||
if s.startswith("```"):
|
||||
first_nl = s.find("\n")
|
||||
if first_nl != -1:
|
||||
s = s[first_nl + 1:]
|
||||
if s.endswith("```"):
|
||||
s = s[:-3]
|
||||
s = s.strip()
|
||||
return s
|
||||
|
||||
|
||||
def _parse_action(text: str) -> Optional[Dict[str, Any]]:
|
||||
s = _strip_fences(text)
|
||||
# First try direct parse, then balanced-brace extraction.
|
||||
try:
|
||||
obj = json.loads(s)
|
||||
if isinstance(obj, dict) and "action" in obj:
|
||||
return obj
|
||||
except json.JSONDecodeError:
|
||||
pass
|
||||
start = s.find("{")
|
||||
if start == -1:
|
||||
return None
|
||||
depth = 0
|
||||
for i in range(start, len(s)):
|
||||
c = s[i]
|
||||
if c == "{":
|
||||
depth += 1
|
||||
elif c == "}":
|
||||
depth -= 1
|
||||
if depth == 0:
|
||||
try:
|
||||
obj = json.loads(s[start : i + 1])
|
||||
if isinstance(obj, dict) and "action" in obj:
|
||||
return obj
|
||||
except json.JSONDecodeError:
|
||||
return None
|
||||
return None
|
||||
|
||||
|
||||
def _extract_final_answer_text(text: str) -> str:
|
||||
"""Best-effort: pull the answer string from a malformed action emission.
|
||||
|
||||
Tries `"answer": "..."` regex, then the GAIA-style `FINAL ANSWER:` line.
|
||||
"""
|
||||
m = re.search(r'"answer"\s*:\s*"((?:\\.|[^"\\])*)"', text, re.DOTALL)
|
||||
if m:
|
||||
return m.group(1).encode("utf-8").decode("unicode_escape")
|
||||
m = re.search(r"FINAL\s*ANSWER\s*:\s*(.+?)\s*$", text, re.IGNORECASE | re.MULTILINE)
|
||||
if m:
|
||||
return m.group(1).strip()
|
||||
return text.strip()
|
||||
@@ -0,0 +1,106 @@
|
||||
"""Prompt strings + tool specs for ToolOrchestraAgent (split from toolorchestra.py)."""
|
||||
|
||||
from __future__ import annotations
|
||||
|
||||
from typing import Any, Dict, List
|
||||
|
||||
ORCHESTRATOR_SYS = """\
|
||||
You are a tool-orchestrating agent. You coordinate a pool of workers to answer the user's question. Each turn you MUST emit exactly one JSON object — no prose, no markdown fences — taking one of two forms:
|
||||
|
||||
{"action": "call_worker", "worker_id": <int>, "input": "<question or instruction for that worker>"}
|
||||
|
||||
{"action": "final_answer", "answer": "<final answer to the user, respecting the question's answer-format rules>"}
|
||||
|
||||
Strategy:
|
||||
|
||||
- Call cheap / specialized workers first (small local model for extraction or arithmetic on given data; web_search for unknowns; specialist LLMs for code/math).
|
||||
- Call the frontier worker (Opus / GPT-5) sparingly, for hard reasoning or a final synthesis pass.
|
||||
- Stop and emit `final_answer` as soon as the previous worker output is sufficient. Do NOT call a worker just to paraphrase.
|
||||
- The user only sees the `answer` field of `final_answer`, so make sure it follows any answer-format rules in the question.
|
||||
"""
|
||||
|
||||
FORCE_FINAL_PROMPT = (
|
||||
"Worker-call budget exhausted. Emit `final_answer` now using everything "
|
||||
"you've learned. Respect the question's answer-format rules."
|
||||
)
|
||||
|
||||
|
||||
# ============================================================================
|
||||
# RL-mode constants (Orchestrator-8B, paper-faithful).
|
||||
# ============================================================================
|
||||
#
|
||||
# Verbatim copies of the upstream system prompt / user-prompt template / tools
|
||||
# from `external/ToolOrchestra/evaluation/eval_hle.py` + `tools.json`. Don't
|
||||
# edit the description text — Orchestrator-8B was RL-trained against this
|
||||
# exact wording and pricing/latency table.
|
||||
|
||||
RL_ORCHESTRATOR_SYS = "You are good at using tools."
|
||||
|
||||
RL_TOOLS_SPEC: List[Dict[str, Any]] = [
|
||||
{
|
||||
"type": "function",
|
||||
"function": {
|
||||
"name": "enhance_reasoning",
|
||||
"description": "tool to enhance answer model reasoning. analyze the problem, write code, execute it and return intermidiate results that will help solve the problem",
|
||||
"parameters": {
|
||||
"properties": {
|
||||
"model": {
|
||||
"description": "The model used to reason. Choices: ['reasoner-1', 'reasoner-2', 'reasoner-3']. reasoner-1 demonstrates strong understanding and reasoning capabilities, which usually provides reliable insights. reasoner-2 can analyze some problems, but could hallucinate and make mistakes in difficult scenarios. reasoner-3 can reason over the context and reveal the logic. \nModel | price per million input tokens | price per million output tokens | average latency\nreasoner-1 | $1.25 | $10 | 31s\nreasoner-2 | $0.25 | $2 | 25s\nreasoner-3 | $0.8 | $0.8 | 9s",
|
||||
"type": "string",
|
||||
}
|
||||
},
|
||||
"required": ["model"],
|
||||
"title": "parameters",
|
||||
"type": "object",
|
||||
},
|
||||
},
|
||||
},
|
||||
{
|
||||
"type": "function",
|
||||
"function": {
|
||||
"name": "answer",
|
||||
"description": "give the final answer. Not allowed to call if documents is empty.",
|
||||
"parameters": {
|
||||
"properties": {
|
||||
"model": {
|
||||
"description": "The model used to answer. Choices: ['answer-1', 'answer-2', 'answer-3', 'answer-4', 'answer-math-1', 'answer-math-2']. answer-1 exhibits strong functional calling abilities and performs excellent in most domains (math, physics, social science, etc.). answer-2 presents reasonable solutions in some tasks, but could get stuck in complex reasoning and specific domain knowledge. answer-3 could solve easy to medium tasks, but is not capable of tackling tasks with strong expertise and long-horizon planning. answer-4 demonstrates basic capability: it can understand basic instructions, do simple steps, yet it sometimes misreads details, mixes concepts. answer-math-1 can solve moderate (middle school) math problem, though it becomes incapable in more difficult tasks. answer-math-2 can follow simple instructions and perform easy (primary-level) math problems, but struggle in more complex logic. The table below shows the pricing and latency of each model:\nModel | price per million input tokens | price per million output tokens | average latency\nanswer-1 | $1.25 | $10 | 96s\nanswer-2 | $0.25 | $2 | 27s\nanswer-3 | $0.9 | $0.9 | 15s\nanswer-4 | $0.8 | $0.8 | 11s\nanswer-math-1 | $0.9 | $0.9 | 13s\nanswer-math-2 | $$0.2 | $0.2 | 9s",
|
||||
"type": "string",
|
||||
}
|
||||
},
|
||||
"required": ["model"],
|
||||
"title": "parameters",
|
||||
"type": "object",
|
||||
},
|
||||
},
|
||||
},
|
||||
{
|
||||
"type": "function",
|
||||
"function": {
|
||||
"name": "search",
|
||||
"description": "Search for missing information",
|
||||
"parameters": {
|
||||
"properties": {
|
||||
"model": {
|
||||
"description": "The model used to search for missing information. Choices: ['search-1', 'search-2', 'search-3']. search-1 usually identifies the missing information and can write concise queries for effective search. search-2 can reason over the context and write queries to find the missing content for answering questions. search-3 can also write queries to find information. The table below shows the pricing and latency:\nModel | price per million input tokens | price per million output tokens | average latency\nsearch-1 | $1.25 | $10 | 22s\nsearch-2 | $0.25 | $2 | 16s\nsearch-3 | $0.8 | $0.8 | 8s",
|
||||
"type": "string",
|
||||
}
|
||||
},
|
||||
"required": ["model"],
|
||||
"title": "parameters",
|
||||
"type": "object",
|
||||
},
|
||||
},
|
||||
},
|
||||
]
|
||||
|
||||
# RL_ALL_TOOLS: argument-validation schema (mirrors eval_hle.py:104).
|
||||
RL_ALL_TOOLS: Dict[str, Dict[str, List[str]]] = {
|
||||
"enhance_reasoning": {"model": ["reasoner-1", "reasoner-2", "reasoner-3"]},
|
||||
"answer": {
|
||||
"model": [
|
||||
"answer-1", "answer-2", "answer-3", "answer-4",
|
||||
"answer-math-1", "answer-math-2",
|
||||
],
|
||||
},
|
||||
"search": {"model": ["search-1", "search-2", "search-3"]},
|
||||
}
|
||||
@@ -0,0 +1,324 @@
|
||||
"""Faithful unified-tool rollout loop for ToolOrchestra (arXiv:2511.21689 §2.2).
|
||||
|
||||
One reasoning->action->observation loop where the orchestrator picks **a named
|
||||
tool** (one per model, from :mod:`expert_registry`) each turn, the environment
|
||||
executes it, and the observation is appended to a running context. The rollout
|
||||
ends when the orchestrator emits a turn with **no tool call** (its text is the
|
||||
final answer) or ``max_turns`` is hit.
|
||||
|
||||
The loop is parameterized over two injected callables so it is pure control flow
|
||||
(no network) and unit-testable with fakes — the agent supplies real ones:
|
||||
|
||||
* ``call_orchestrator(messages, tool_specs) -> (text, tool_calls, p_tok, c_tok)``
|
||||
where ``messages`` is the running system/user/assistant/tool conversation and
|
||||
``tool_calls`` is a list of ``(name, arguments)`` (possibly empty).
|
||||
* ``dispatch(tool, arguments) -> (observation, cost_usd, tokens, is_local)``.
|
||||
"""
|
||||
|
||||
from __future__ import annotations
|
||||
|
||||
import json
|
||||
import random
|
||||
from dataclasses import dataclass, field
|
||||
from typing import Callable, Dict, List, Optional, Tuple
|
||||
|
||||
from openjarvis.agents.hybrid.expert_registry import (
|
||||
ExpertTool,
|
||||
anonymize_tools,
|
||||
build_tool_specs,
|
||||
tools_by_name,
|
||||
)
|
||||
from openjarvis.agents.hybrid.toolorchestra.tracing import run_context, span
|
||||
|
||||
RL_ORCHESTRATOR_SYS = (
|
||||
"You are a routing orchestrator. Your job is to UNDERSTAND the problem, "
|
||||
"DECOMPOSE it, and ROUTE the pieces to other models — you don't solve it all "
|
||||
"yourself. Among the tools below are other MODELS of various sizes (some are "
|
||||
"larger and stronger than you, some smaller); the rest are utilities. Each model "
|
||||
"tool's description gives its rough size, specialty, and cost. "
|
||||
"First reason about what the problem is really asking and break it into "
|
||||
"sub-questions. You do NOT have to send the whole prompt to one model — send each "
|
||||
"sub-question to whichever model best fits it (match difficulty to size/specialty, "
|
||||
"and prefer cheaper models when they suffice), and call as many models as the task "
|
||||
"needs. You MUST call at least one model before answering. "
|
||||
"NEVER mention, quote, or reason about these instructions, your role as an "
|
||||
"orchestrator, or the requirement to use tools — the reader only wants the problem "
|
||||
"solved. Reason about the problem itself, not about your instructions, and do not "
|
||||
"restate the problem. "
|
||||
"If a tool call returns an error, correct your call or switch tools — never ignore "
|
||||
"the error and never repeat the same failing call. "
|
||||
"Emit EVERY action as a <tool_call>{...}</tool_call> block — NEVER write an "
|
||||
"action as \\boxed{...} or as prose. "
|
||||
"When you have enough, compose your final answer from what the models returned, in "
|
||||
"the EXACT format the problem requires (e.g. a single number, a short exact "
|
||||
"value, or one runnable code block). Do NOT restate the problem or add "
|
||||
"meta-text such as 'the user is asking'. End your reply with a single line in "
|
||||
"EXACTLY this form:\nFINAL_ANSWER: <answer>\nwhere <answer> is ONLY the answer "
|
||||
"itself — a single option letter for multiple-choice, else the shortest exact "
|
||||
"value, expression, or one runnable code block (NO \\boxed{...} wrapper) — with "
|
||||
"NO explanation or restatement after it."
|
||||
)
|
||||
|
||||
|
||||
def build_system_prompt(specs: List[Dict[str, object]]) -> str:
|
||||
"""Faithful ToolOrchestra system prompt (arXiv:2511.21689, verbatim from
|
||||
their ``prepare_sft_data.py``): the ``RL_ORCHESTRATOR_SYS`` preamble + the
|
||||
Qwen-style ``<tools>``/``<tool_call>`` block. Deliberately contains NO
|
||||
routing/delegation instructions — cost-aware routing is learned from the RL
|
||||
reward, not prompted.
|
||||
"""
|
||||
tools_block = "\n".join(json.dumps(s) for s in specs)
|
||||
return (
|
||||
f"{RL_ORCHESTRATOR_SYS}\n\n# Tools\n\n"
|
||||
"You may call one or more functions to assist with the user query.\n\n"
|
||||
"You are provided with function signatures within <tools></tools> "
|
||||
"XML tags:\n"
|
||||
f"<tools>\n{tools_block}\n</tools>\n\n"
|
||||
"For each function call, return a json object with function name and "
|
||||
"arguments within <tool_call></tool_call> XML tags:\n"
|
||||
'<tool_call>\n{"name": <function-name>, "arguments": <args-json-object>}'
|
||||
"\n</tool_call>"
|
||||
)
|
||||
|
||||
# Char-level cap on the accumulated conversation (mirrors the paper's ~24k-token
|
||||
# cap) and a per-observation cap so one giant tool dump can't blow it out.
|
||||
_CONTEXT_CAP = 24000
|
||||
_OBS_CAP = 8000
|
||||
# After this many tool calls, push the orchestrator to stop and answer — bounds
|
||||
# the "keep re-asking the same model forever" loop (esp. gemma).
|
||||
_SOFT_CALL_CAP = 6
|
||||
|
||||
|
||||
def _trim_history(messages: List[Dict[str, str]]) -> None:
|
||||
"""Keep the message history under ``_CONTEXT_CAP`` chars by dropping the
|
||||
OLDEST assistant/tool exchange, never the system prompt or the problem."""
|
||||
def total() -> int:
|
||||
return sum(len(m.get("content") or "") for m in messages)
|
||||
# messages[0]=system, messages[1]=user problem — always keep those two.
|
||||
while total() > _CONTEXT_CAP and len(messages) > 4:
|
||||
del messages[2:4]
|
||||
|
||||
|
||||
@dataclass
|
||||
class UnifiedTurn:
|
||||
"""One orchestrator turn. ``tool_name is None`` marks the final-answer turn."""
|
||||
|
||||
reasoning: str
|
||||
tool_name: Optional[str] = None
|
||||
arguments: Dict[str, object] = field(default_factory=dict)
|
||||
observation: Optional[str] = None
|
||||
|
||||
|
||||
@dataclass
|
||||
class UnifiedRollout:
|
||||
turns: List[UnifiedTurn]
|
||||
final_answer: str
|
||||
cost_usd: float = 0.0
|
||||
tokens: int = 0
|
||||
num_tool_calls: int = 0
|
||||
parse_failures: int = 0
|
||||
# When experts were anonymized for this rollout, maps the opaque label the
|
||||
# policy saw (e.g. ``expert_a3f9``) back to the real tool name (``gpt_5_5``).
|
||||
anon_map: Optional[Dict[str, str]] = None
|
||||
# The EXACT tool specs the policy saw this rollout (anonymized when
|
||||
# ``anonymize=True``). Serialization MUST build the saved system prompt from
|
||||
# these — not from the real registry — or the prompt's ``<tools>`` block ends
|
||||
# up with real model names+pricing while the assistant ``<tool_call>`` tags
|
||||
# use the anon labels, which both breaks SFT (calls a tool absent from its
|
||||
# list) and re-injects the name/cost bias anonymization removed.
|
||||
tool_specs: Optional[List[Dict[str, object]]] = None
|
||||
|
||||
def tool_calls(self) -> List[Tuple[str, Dict[str, object]]]:
|
||||
return [(t.tool_name, t.arguments) for t in self.turns if t.tool_name]
|
||||
|
||||
|
||||
def _tool_prompt(tool: ExpertTool, arguments: Dict[str, object], question: str) -> str:
|
||||
"""The text we actually send the dispatched tool, framed by its arg schema."""
|
||||
for key in ("input", "query", "code"):
|
||||
val = arguments.get(key)
|
||||
if isinstance(val, str) and val.strip():
|
||||
return val
|
||||
return question
|
||||
|
||||
|
||||
def _run_unified_rollout_inner(
|
||||
question: str,
|
||||
tools: List[ExpertTool],
|
||||
*,
|
||||
call_orchestrator: Callable[..., Tuple[str, List[Tuple[str, Dict[str, object]]], int, int]],
|
||||
dispatch: Callable[[ExpertTool, Dict[str, object]], Tuple[str, float, int, bool]],
|
||||
max_turns: int = 50,
|
||||
system: str = RL_ORCHESTRATOR_SYS,
|
||||
anonymize: bool = False,
|
||||
) -> UnifiedRollout:
|
||||
"""Drive the faithful unified-tool rollout for one task.
|
||||
|
||||
``anonymize``: replace each model expert with an opaque random label, a
|
||||
uniform description and no cost line, shuffled — so the policy can't route on
|
||||
a model's name/position/cost (which we found dominate the choice). The
|
||||
anon->real mapping is returned on the rollout for offline analysis.
|
||||
"""
|
||||
anon_map: Optional[Dict[str, str]] = None
|
||||
if anonymize:
|
||||
tools, anon_map = anonymize_tools(tools, random.Random())
|
||||
specs = build_tool_specs(tools)
|
||||
by_name = tools_by_name(tools)
|
||||
|
||||
# Proper multi-turn conversation (NOT a flattened user blob): the orchestrator
|
||||
# sees its own calls as `assistant` turns and each observation as a `tool`
|
||||
# turn, so it can tell a tool result from user input and doesn't re-derive the
|
||||
# whole problem every turn. This mirrors the serialized SFT format exactly.
|
||||
messages: List[Dict[str, str]] = [
|
||||
{"role": "system", "content": system},
|
||||
{"role": "user", "content": f"Problem: {question}"},
|
||||
]
|
||||
turns: List[UnifiedTurn] = []
|
||||
cost = 0.0
|
||||
tokens = 0
|
||||
n_tool_calls = 0
|
||||
parse_failures = 0
|
||||
nudges = 0
|
||||
empty_input_nudges = 0
|
||||
final_answer = ""
|
||||
|
||||
for _ in range(max_turns):
|
||||
text, tool_calls, p_tok, c_tok = call_orchestrator(messages, specs)
|
||||
tokens += int(p_tok) + int(c_tok)
|
||||
|
||||
if not tool_calls:
|
||||
# Enforce the "MUST delegate to a model" rule: if the orchestrator
|
||||
# tries to answer before ANY tool call, nudge it to route instead of
|
||||
# accepting a solve-it-yourself answer (which reject-sampling drops).
|
||||
if n_tool_calls == 0 and nudges < 2:
|
||||
nudges += 1
|
||||
messages.append({"role": "assistant", "content": text or ""})
|
||||
messages.append({"role": "user", "content": (
|
||||
"Make progress by delegating a concrete sub-question to one of the "
|
||||
"models now via a <tool_call>.")})
|
||||
continue
|
||||
# No tool call -> the orchestrator is answering. Terminate.
|
||||
final_answer = (text or "").strip()
|
||||
turns.append(UnifiedTurn(reasoning=text or "", tool_name=None))
|
||||
messages.append({"role": "assistant", "content": text or ""})
|
||||
break
|
||||
|
||||
name, arguments = tool_calls[0]
|
||||
if name not in by_name:
|
||||
parse_failures += 1
|
||||
messages.append({"role": "assistant", "content": text or ""})
|
||||
messages.append({"role": "user",
|
||||
"content": f"[invalid tool {name!r} — choose one from the provided tool list]"})
|
||||
if parse_failures >= 2:
|
||||
final_answer = (text or "").strip()
|
||||
break
|
||||
continue
|
||||
|
||||
tool = by_name[name]
|
||||
# Empty-input guard: the small orchestrator often emits a <tool_call> with
|
||||
# a blank/missing input on harder multi-turn tasks. Dispatching it returns
|
||||
# a "no input provided" error observation that poisons the whole (often
|
||||
# otherwise-correct) trajectory — the dominant clean-yield killer on hard
|
||||
# tasks. Instead, nudge the model to resend WITH input and drop the
|
||||
# malformed turn (don't record it), so it never enters the training data.
|
||||
if tool.kind == "model" or tool.backend_type != "openjarvis-tool":
|
||||
_has_input = any(
|
||||
isinstance(arguments.get(k), str) and arguments.get(k).strip()
|
||||
for k in ("input", "query", "code")
|
||||
)
|
||||
if not _has_input and empty_input_nudges < 3:
|
||||
empty_input_nudges += 1
|
||||
messages.append({"role": "assistant", "content": text or ""})
|
||||
messages.append({"role": "user", "content": (
|
||||
f"Your call to {name} had an empty 'input'. Resend the "
|
||||
"<tool_call> with a non-empty 'input' field containing the "
|
||||
"concrete sub-question to delegate.")})
|
||||
continue
|
||||
obs, dcost, dtok, _is_local = dispatch(tool, arguments)
|
||||
cost += float(dcost)
|
||||
tokens += int(dtok)
|
||||
n_tool_calls += 1
|
||||
turns.append(UnifiedTurn(
|
||||
reasoning=text or "", tool_name=name, arguments=dict(arguments),
|
||||
observation=obs,
|
||||
))
|
||||
# The model's own action as an assistant turn (the <tool_call> tag in
|
||||
# content, matching the SFT serialization), then the observation as a
|
||||
# distinct `tool` turn the model reads as a tool response.
|
||||
call_content = text or ""
|
||||
if "<tool_call>" not in call_content:
|
||||
call_content = (call_content + "\n" + tool_call_tag(name, arguments)).strip()
|
||||
messages.append({"role": "assistant", "content": call_content})
|
||||
obs_text = obs or ""
|
||||
if len(obs_text) > _OBS_CAP:
|
||||
obs_text = obs_text[:_OBS_CAP] + "\n…[truncated]"
|
||||
messages.append({"role": "tool", "name": name, "content": obs_text})
|
||||
if n_tool_calls >= _SOFT_CALL_CAP:
|
||||
messages.append({"role": "user", "content": (
|
||||
"You now have enough information from the models. Do NOT call any "
|
||||
"more tools — reply with your FINAL_ANSWER line only.")})
|
||||
_trim_history(messages)
|
||||
else:
|
||||
# Hit max_turns with no explicit answer: use the last observation/text.
|
||||
final_answer = (turns[-1].observation or turns[-1].reasoning).strip() if turns else ""
|
||||
|
||||
return UnifiedRollout(
|
||||
turns=turns, final_answer=final_answer, cost_usd=cost, tokens=tokens,
|
||||
num_tool_calls=n_tool_calls, parse_failures=parse_failures,
|
||||
anon_map=anon_map, tool_specs=specs,
|
||||
)
|
||||
|
||||
|
||||
def run_unified_rollout(
|
||||
question: str,
|
||||
tools: List[ExpertTool],
|
||||
*,
|
||||
call_orchestrator: Callable[..., Tuple[str, List[Tuple[str, Dict[str, object]]], int, int]],
|
||||
dispatch: Callable[[ExpertTool, Dict[str, object]], Tuple[str, float, int, bool]],
|
||||
max_turns: int = 50,
|
||||
system: str = RL_ORCHESTRATOR_SYS,
|
||||
anonymize: bool = False,
|
||||
) -> UnifiedRollout:
|
||||
"""One ToolOrchestra rollout = one Braintrust trace. Opens the root span,
|
||||
runs the loop (orchestrator turns + expert/tool dispatches nest inside as
|
||||
child spans), then logs the final answer + cost/tokens. No-op when tracing
|
||||
is disabled (see :mod:`.tracing`)."""
|
||||
_run_meta, _run_tags = run_context()
|
||||
_fields = dict(
|
||||
input={"question": question},
|
||||
metadata={"max_turns": max_turns, "anonymize": anonymize,
|
||||
"n_tools": len(tools), **_run_meta},
|
||||
)
|
||||
if _run_tags:
|
||||
_fields["tags"] = _run_tags
|
||||
with span("toolorchestra.rollout", span_type="task", **_fields) as _s:
|
||||
roll = _run_unified_rollout_inner(
|
||||
question, tools, call_orchestrator=call_orchestrator, dispatch=dispatch,
|
||||
max_turns=max_turns, system=system, anonymize=anonymize,
|
||||
)
|
||||
_s.log(
|
||||
output=roll.final_answer,
|
||||
metrics={"cost_usd": roll.cost_usd, "tokens": roll.tokens,
|
||||
"num_tool_calls": roll.num_tool_calls,
|
||||
"parse_failures": roll.parse_failures},
|
||||
metadata={"n_experts_available": len(roll.anon_map or {}),
|
||||
"answered": bool(roll.final_answer),
|
||||
# label -> real model, so every anonymized route span in this
|
||||
# trace can be decoded back to the model that actually ran.
|
||||
"anon_map": roll.anon_map or {}},
|
||||
)
|
||||
return roll
|
||||
|
||||
|
||||
def tool_call_tag(name: str, arguments: Dict[str, object]) -> str:
|
||||
"""Render a tool call as the ``<tool_call>{...}</tool_call>`` text the model emits."""
|
||||
return f"<tool_call>{json.dumps({'name': name, 'arguments': arguments})}</tool_call>"
|
||||
|
||||
|
||||
__all__ = [
|
||||
"RL_ORCHESTRATOR_SYS",
|
||||
"UnifiedRollout",
|
||||
"UnifiedTurn",
|
||||
"build_system_prompt",
|
||||
"run_unified_rollout",
|
||||
"tool_call_tag",
|
||||
]
|
||||
@@ -0,0 +1,79 @@
|
||||
"""Tavily search + Modal Python sandbox helpers for ToolOrchestraAgent."""
|
||||
|
||||
from __future__ import annotations
|
||||
|
||||
import re
|
||||
from typing import Optional, Tuple
|
||||
|
||||
# ---- Tavily + Modal helpers -------------------------------------------------
|
||||
|
||||
def _call_tavily_search(query: str, max_results: int = 5) -> Tuple[str, int, int]:
|
||||
"""One-shot Tavily search. Returns (text, p_tok=0, c_tok=0).
|
||||
|
||||
Token counts are reported as zero (no LLM was billed); the OpenJarvis
|
||||
accounting layer separately tallies tool-call counts. Falls back to
|
||||
DuckDuckGo if Tavily is unreachable (see ``WebSearchTool``).
|
||||
"""
|
||||
from openjarvis.tools.web_search import WebSearchTool
|
||||
|
||||
tool = WebSearchTool(max_results=max_results)
|
||||
res = tool.execute(query=query, max_results=max_results)
|
||||
text = res.content or ""
|
||||
if not res.success and not text:
|
||||
text = "(no results)"
|
||||
return text, 0, 0
|
||||
|
||||
|
||||
_MODAL_APP_NAME = "openjarvis-toolorchestra-sandbox"
|
||||
|
||||
|
||||
def _call_modal_python(code: str, timeout_s: int = 60) -> Tuple[str, int]:
|
||||
"""Execute a single Python snippet in a fresh Modal Sandbox.
|
||||
|
||||
Returns ``(combined_stdout_stderr, returncode)``. Logs are capped at 8 KiB.
|
||||
Any exception (modal auth, network, sandbox boot failure) is captured into
|
||||
the returned string with a non-zero rc — we never raise back to the
|
||||
orchestrator loop. The sandbox is torn down at the end via ``terminate()``.
|
||||
"""
|
||||
try:
|
||||
import modal
|
||||
|
||||
app = modal.App.lookup(_MODAL_APP_NAME, create_if_missing=True)
|
||||
# python:3.12-slim is small + boots fast; the paper uses a generic
|
||||
# Python image too. We rely on stdlib only — no extra pip installs.
|
||||
image = modal.Image.debian_slim(python_version="3.12")
|
||||
sb = modal.Sandbox.create(
|
||||
"python", "-c", code,
|
||||
app=app,
|
||||
image=image,
|
||||
timeout=int(timeout_s),
|
||||
)
|
||||
sb.wait()
|
||||
try:
|
||||
out = sb.stdout.read() or ""
|
||||
except Exception:
|
||||
out = ""
|
||||
try:
|
||||
err = sb.stderr.read() or ""
|
||||
except Exception:
|
||||
err = ""
|
||||
rc = sb.returncode if sb.returncode is not None else -1
|
||||
try:
|
||||
sb.terminate()
|
||||
except Exception:
|
||||
pass
|
||||
combined = out + (("\n" + err) if err else "")
|
||||
if len(combined) > 8192:
|
||||
combined = combined[:8192] + "\n... (output truncated)"
|
||||
return combined, int(rc)
|
||||
except Exception as exc:
|
||||
return f"[modal-python error: {type(exc).__name__}: {exc}]", -1
|
||||
|
||||
|
||||
_PY_CODE_RE = re.compile(r"```(?:python|py)?\s*\n(.*?)```", re.DOTALL)
|
||||
|
||||
|
||||
def _extract_first_python_block(text: str) -> Optional[str]:
|
||||
"""Return the first ```python ... ``` block (or ```...```), or None."""
|
||||
m = _PY_CODE_RE.search(text or "")
|
||||
return m.group(1).strip() if m else None
|
||||
@@ -0,0 +1,161 @@
|
||||
"""Braintrust telemetry for the ToolOrchestra rollout.
|
||||
|
||||
Each ``run_unified_rollout`` becomes ONE Braintrust trace (root span); the
|
||||
orchestrator turns and every expert/tool dispatch nest inside it, and the
|
||||
underlying OpenAI/Anthropic calls (wrapped clients) nest one level deeper — so
|
||||
you see the full routing tree with inputs/outputs/tokens/cost per node.
|
||||
|
||||
On by default. Degrades to a total no-op (never raises, never changes behavior)
|
||||
when: ``OJ_BRAINTRUST=0``, the ``braintrust`` package isn't installed, or
|
||||
``BRAINTRUST_API_KEY`` is unset. Concurrency: rollouts run one-per-thread
|
||||
(ThreadPoolExecutor); threads don't inherit contextvars, so each rollout opens
|
||||
its own independent root trace and same-thread child calls nest correctly.
|
||||
"""
|
||||
|
||||
from __future__ import annotations
|
||||
|
||||
import contextlib
|
||||
import logging
|
||||
import os
|
||||
from typing import Any, Optional
|
||||
|
||||
logger = logging.getLogger(__name__)
|
||||
|
||||
_STATE: dict = {"resolved": False, "enabled": False, "bt": None}
|
||||
|
||||
|
||||
def _truthy(v: str) -> bool:
|
||||
return v.strip().lower() not in ("0", "false", "no", "off", "")
|
||||
|
||||
|
||||
def _resolve() -> bool:
|
||||
"""Lazily decide whether tracing is active and init the logger once."""
|
||||
if _STATE["resolved"]:
|
||||
return _STATE["enabled"]
|
||||
_STATE["resolved"] = True
|
||||
if not _truthy(os.getenv("OJ_BRAINTRUST", "1")): # on by default
|
||||
return False
|
||||
if not os.getenv("BRAINTRUST_API_KEY"):
|
||||
logger.info("braintrust on-by-default but BRAINTRUST_API_KEY unset — tracing disabled")
|
||||
return False
|
||||
try:
|
||||
import braintrust as _bt
|
||||
|
||||
proj_id = os.getenv("OJ_BRAINTRUST_PROJECT_ID")
|
||||
if proj_id:
|
||||
_bt.init_logger(project_id=proj_id)
|
||||
else:
|
||||
_bt.init_logger(project=os.getenv("OJ_BRAINTRUST_PROJECT", "toolorchestra"))
|
||||
_STATE["bt"] = _bt
|
||||
_STATE["enabled"] = True
|
||||
logger.info("braintrust tracing ENABLED (%s)",
|
||||
f"project_id={proj_id}" if proj_id
|
||||
else f"project={os.getenv('OJ_BRAINTRUST_PROJECT', 'toolorchestra')}")
|
||||
except Exception as exc: # missing pkg / bad key / init failure — never crash
|
||||
logger.warning("braintrust init failed (%s) — tracing disabled", exc)
|
||||
return _STATE["enabled"]
|
||||
|
||||
|
||||
def enabled() -> bool:
|
||||
return _resolve()
|
||||
|
||||
|
||||
def run_context() -> tuple[dict, list]:
|
||||
"""Run-level metadata + tags for the ROOT rollout span, sourced from env
|
||||
(set by the generation driver). No-op-safe: returns ({}, []) if nothing is
|
||||
set, and never raises. Env keys:
|
||||
|
||||
- ``OJ_RUN_LABEL`` — human run label (also stamped on the uploaded dataset).
|
||||
- ``OJ_GEN_MODEL`` — SPECIFIC gen model id (e.g. ``Qwen/Qwen3.5-9B``).
|
||||
- ``OJ_RUN_STAGE`` — ``prod``|``smoke`` (inferred from the label if unset).
|
||||
- ``OJ_CFG_*`` — config knobs (temperature/max_turns/anonymize/...).
|
||||
"""
|
||||
import datetime
|
||||
|
||||
meta: dict = {}
|
||||
tags: list = []
|
||||
try:
|
||||
label = os.getenv("OJ_RUN_LABEL")
|
||||
gen_model = os.getenv("OJ_GEN_MODEL")
|
||||
stage = os.getenv("OJ_RUN_STAGE")
|
||||
if not stage and label:
|
||||
stage = "smoke" if "smoke" in label.lower() else "prod"
|
||||
if label:
|
||||
meta["run_label"] = label
|
||||
if gen_model:
|
||||
meta["gen_model"] = gen_model
|
||||
if stage:
|
||||
meta["stage"] = stage
|
||||
cfg = {}
|
||||
for env_key, key in (("OJ_CFG_TEMPERATURE", "temperature"),
|
||||
("OJ_CFG_MAX_TURNS", "max_turns"),
|
||||
("OJ_CFG_ANONYMIZE", "anonymize"),
|
||||
("OJ_CFG_REJECTION_ONLY", "rejection_only")):
|
||||
v = os.getenv(env_key)
|
||||
if v not in (None, ""):
|
||||
cfg[key] = v
|
||||
if cfg:
|
||||
meta["config"] = cfg
|
||||
date = datetime.date.today().isoformat()
|
||||
tags = [t for t in (gen_model, stage, date) if t]
|
||||
except Exception: # never let telemetry enrichment break a rollout
|
||||
return {}, []
|
||||
return meta, tags
|
||||
|
||||
|
||||
def wrap_client(client: Any) -> Any:
|
||||
"""Wrap an OpenAI/Anthropic client so its calls auto-log under the current
|
||||
span. Pass-through (unchanged client) when tracing is off or on any error —
|
||||
so this is always safe to call at client construction."""
|
||||
if not _resolve():
|
||||
return client
|
||||
try:
|
||||
bt = _STATE["bt"]
|
||||
mod = type(client).__module__.lower()
|
||||
if "anthropic" in mod:
|
||||
return bt.wrap_anthropic(client)
|
||||
return bt.wrap_openai(client)
|
||||
except Exception as exc:
|
||||
logger.warning("braintrust wrap_client failed (%s) — using raw client", exc)
|
||||
return client
|
||||
|
||||
|
||||
class _NullSpan:
|
||||
"""No-op span used when tracing is disabled (keeps call sites branch-free)."""
|
||||
|
||||
def log(self, **_kw: Any) -> None:
|
||||
pass
|
||||
|
||||
def __enter__(self) -> "_NullSpan":
|
||||
return self
|
||||
|
||||
def __exit__(self, *_a: Any) -> bool:
|
||||
return False
|
||||
|
||||
|
||||
@contextlib.contextmanager
|
||||
def span(name: str, *, span_type: str = "task", **fields: Any):
|
||||
"""Open a Braintrust span (or a no-op). Use as::
|
||||
|
||||
with span("toolorchestra.rollout", input=...) as s:
|
||||
...
|
||||
s.log(output=..., metrics=..., metadata=...)
|
||||
|
||||
``fields`` (input/metadata/...) are forwarded to ``start_span``.
|
||||
"""
|
||||
if not _resolve():
|
||||
yield _NullSpan()
|
||||
return
|
||||
# Guard only span CREATION (so a Braintrust hiccup degrades to a no-op). Do
|
||||
# NOT wrap the yielded body in try/except: an exception from the rollout would
|
||||
# be thrown into this generator, caught here, and masked by a second yield
|
||||
# ("generator didn't stop after throw()"). Let the body's own exceptions
|
||||
# propagate through the `with` untouched — Braintrust records + re-raises.
|
||||
try:
|
||||
cm = _STATE["bt"].start_span(name=name, type=span_type, **fields)
|
||||
except Exception as exc: # span creation failed — run trace-less, never break
|
||||
logger.warning("braintrust start_span(%s) failed (%s) — continuing untraced", name, exc)
|
||||
yield _NullSpan()
|
||||
return
|
||||
with cm as s:
|
||||
yield s
|
||||
@@ -0,0 +1,255 @@
|
||||
"""Real backends for the unified-tool rollout — the bridge between the pure
|
||||
:func:`run_unified_rollout` loop and live model/tool calls.
|
||||
|
||||
``make_call_orchestrator`` returns the ``call_orchestrator`` callable (a teacher
|
||||
LLM emitting tool calls over the unified spec); ``make_dispatch`` returns the
|
||||
``dispatch`` callable (executes a chosen tool via ``_call_worker``). Both are
|
||||
import-safe — the OpenAI SDK is imported lazily so this module loads without
|
||||
network or keys.
|
||||
"""
|
||||
|
||||
from __future__ import annotations
|
||||
|
||||
import json
|
||||
import threading
|
||||
from typing import Any, Callable, Dict, List, Optional, Tuple
|
||||
|
||||
# Guards the lazy, one-time build of the shared ToolExecutor in
|
||||
# ``_dispatch_openjarvis_tool``: with concurrent rollouts (parallel rejection
|
||||
# sampling) several threads can reach the build at once and would each instantiate
|
||||
# the full tool registry. The lock makes it build-once.
|
||||
_EXECUTOR_BUILD_LOCK = threading.Lock()
|
||||
|
||||
from openjarvis.agents.hybrid._prices import cost as _model_cost
|
||||
from openjarvis.agents.hybrid.expert_registry import ExpertTool, to_worker_dict
|
||||
from openjarvis.agents.hybrid.toolorchestra.parsing import _parse_rl_tool_call
|
||||
from openjarvis.agents.hybrid.toolorchestra.rollout import (
|
||||
UnifiedRollout,
|
||||
build_system_prompt,
|
||||
run_unified_rollout,
|
||||
)
|
||||
from openjarvis.agents.hybrid.toolorchestra.workers import _call_worker
|
||||
from openjarvis.agents.hybrid.toolorchestra.tracing import span
|
||||
|
||||
|
||||
def make_call_orchestrator(
|
||||
model: str,
|
||||
*,
|
||||
base_url: Optional[str] = None,
|
||||
api_key: Optional[str] = None,
|
||||
temperature: float = 1.0,
|
||||
max_tokens: int = 4096,
|
||||
timeout: float = 600.0,
|
||||
native_tools: bool = True,
|
||||
frequency_penalty: float = 0.2,
|
||||
presence_penalty: float = 0.0,
|
||||
repetition_penalty: float = 1.15,
|
||||
) -> Callable[..., Tuple[str, List[Tuple[str, Dict[str, Any]]], int, int]]:
|
||||
"""Teacher-orchestrator caller. ``base_url=None`` → OpenAI cloud; set it to a
|
||||
vLLM endpoint (with ``api_key="EMPTY"``) to drive a local served teacher.
|
||||
|
||||
``native_tools`` picks the tool-calling convention, and the two modes MUST NOT
|
||||
be mixed on the same model:
|
||||
|
||||
* ``True`` (default — DATA GENERATION with a *base* model): pass the OpenAI
|
||||
``tools=`` param so the server's native tool template + parser (gemma4 /
|
||||
qwen3_xml / hermes) drive the base model, which only emits reliable tool
|
||||
calls that way. The rollout captures ``(name, arguments)`` and the serializer
|
||||
re-renders them into the canonical JSON ``<tool_call>`` form regardless.
|
||||
* ``False`` (SERVING a *fine-tuned* model): bake the ``<tools>`` block + JSON
|
||||
``<tool_call>`` format into the system prompt (exactly what the serializer
|
||||
trained on) and DROP ``tools=``. If ``tools=`` is set here, the served chat
|
||||
template injects its own XML ``<function=>`` instructions, which conflict
|
||||
with the JSON format the model learned -> it falls back to ``\\boxed{}`` and
|
||||
never routes. ``_parse_rl_tool_call`` scrapes the JSON tag from raw text.
|
||||
"""
|
||||
|
||||
# Create the client ONCE and reuse it for every turn. Creating a fresh OpenAI
|
||||
# client per call (as before) leaks an httpx connection pool each time -> under
|
||||
# heavy parallel generation, sockets pile up in CLOSE-WAIT, the run degrades and
|
||||
# has to be restarted. The orchestrator is the highest-frequency call, so reusing
|
||||
# one client here removes the dominant leak. httpx clients are thread-safe.
|
||||
from openai import OpenAI
|
||||
|
||||
from openjarvis.agents.hybrid.toolorchestra.tracing import wrap_client
|
||||
|
||||
client = wrap_client(OpenAI(base_url=base_url, api_key=api_key or "EMPTY", timeout=timeout))
|
||||
|
||||
def call_orchestrator(messages: List[Dict[str, Any]], specs: List[Dict[str, Any]]):
|
||||
# ``messages`` is the full running conversation (system/user/assistant/tool)
|
||||
# built by run_unified_rollout — pass it through so the model sees its own
|
||||
# prior calls + tool responses, not a flattened user blob.
|
||||
if native_tools:
|
||||
# Base-model generation: native tool template drives the tool calls.
|
||||
send = messages
|
||||
kwargs = {"tools": specs} if specs else {}
|
||||
else:
|
||||
# Fine-tuned serving: JSON format is baked into the system prompt and
|
||||
# tools= is dropped (see make_call_orchestrator docstring).
|
||||
send = list(messages)
|
||||
if specs and send and send[0].get("role") == "system":
|
||||
send[0] = {"role": "system", "content": build_system_prompt(specs)}
|
||||
kwargs = {}
|
||||
# repetition_penalty is a vLLM extra (not OpenAI-native); only send it to a
|
||||
# local vLLM endpoint (base_url set), never to the cloud frontier APIs.
|
||||
if base_url and repetition_penalty and repetition_penalty != 1.0:
|
||||
kwargs.setdefault("extra_body", {})["repetition_penalty"] = repetition_penalty
|
||||
resp = client.chat.completions.create(
|
||||
model=model,
|
||||
messages=send,
|
||||
temperature=temperature,
|
||||
max_tokens=max_tokens,
|
||||
frequency_penalty=frequency_penalty,
|
||||
presence_penalty=presence_penalty,
|
||||
**kwargs,
|
||||
)
|
||||
msg = resp.choices[0].message
|
||||
text = msg.content or ""
|
||||
sdk_tool_calls = getattr(msg, "tool_calls", None)
|
||||
u = resp.usage
|
||||
p = getattr(u, "prompt_tokens", 0) if u else 0
|
||||
c = getattr(u, "completion_tokens", 0) if u else 0
|
||||
|
||||
parsed = _parse_rl_tool_call(text, sdk_tool_calls)
|
||||
tool_calls = [(parsed["name"], parsed["arguments"])] if parsed else []
|
||||
return text, tool_calls, int(p), int(c)
|
||||
|
||||
return call_orchestrator
|
||||
|
||||
|
||||
def _dispatch_openjarvis_tool(
|
||||
tool: ExpertTool,
|
||||
arguments: Dict[str, Any],
|
||||
executor_holder: Dict[str, Any],
|
||||
) -> Tuple[str, float, int, bool]:
|
||||
"""Execute a bridged real OpenJarvis tool via its ``ToolExecutor``.
|
||||
|
||||
Builds the executor lazily (cached in ``executor_holder``) and degrades to a
|
||||
clear error string — never a crash — if the tool registry / executor can't be
|
||||
instantiated. Returns ``(content, cost_usd, total_tokens, is_local=True)``.
|
||||
"""
|
||||
try:
|
||||
executor = executor_holder.get("executor")
|
||||
if executor is None:
|
||||
with _EXECUTOR_BUILD_LOCK:
|
||||
# Re-check inside the lock: another thread may have built it while
|
||||
# we waited.
|
||||
executor = executor_holder.get("executor")
|
||||
if executor is None:
|
||||
from openjarvis.core.registry import ToolRegistry
|
||||
from openjarvis.tools._stubs import ToolExecutor
|
||||
|
||||
instances = []
|
||||
for name in ToolRegistry.keys():
|
||||
entry = ToolRegistry.get(name)
|
||||
try:
|
||||
instances.append(
|
||||
entry() if isinstance(entry, type) else entry
|
||||
)
|
||||
except Exception:
|
||||
continue
|
||||
# Auto-approve confirmation-gated tools (shell_exec,
|
||||
# git_commit, ...): rollouts/eval run headless in a sandbox, so
|
||||
# there is no human to confirm. Without this, those tools return
|
||||
# a "requires confirmation" error instead of executing — which
|
||||
# would silently break TerminalBench.
|
||||
executor = ToolExecutor(
|
||||
instances,
|
||||
interactive=True,
|
||||
confirm_callback=lambda _prompt: True,
|
||||
)
|
||||
executor_holder["executor"] = executor
|
||||
|
||||
from openjarvis.core.types import ToolCall
|
||||
|
||||
call = ToolCall(
|
||||
id=f"orch-{tool.name}",
|
||||
name=str(tool.model),
|
||||
arguments=json.dumps(arguments or {}),
|
||||
)
|
||||
result = executor.execute(call)
|
||||
usage = getattr(result, "usage", None) or {}
|
||||
total_tokens = int(usage.get("total_tokens", 0) or 0)
|
||||
return (
|
||||
str(getattr(result, "content", "")),
|
||||
float(getattr(result, "cost_usd", 0.0) or 0.0),
|
||||
total_tokens,
|
||||
True,
|
||||
)
|
||||
except Exception as exc: # never crash the rollout on a tool-bridge failure
|
||||
return (f"[openjarvis-tool error: {tool.model}: {exc}]", 0.0, 0, True)
|
||||
|
||||
|
||||
def make_dispatch(
|
||||
cfg: Optional[Dict[str, Any]] = None,
|
||||
) -> Callable[[ExpertTool, Dict[str, Any]], Tuple[str, float, int, bool]]:
|
||||
"""Tool-execution caller: run the chosen tool and return (obs, cost, tokens, is_local)."""
|
||||
cfg = cfg or {}
|
||||
executor_holder: Dict[str, Any] = {}
|
||||
|
||||
def _dispatch_inner(tool: ExpertTool, arguments: Dict[str, Any]):
|
||||
# Bridged real OpenJarvis tools run through the ToolExecutor, not
|
||||
# the model/worker path.
|
||||
if tool.backend_type == "openjarvis-tool":
|
||||
return _dispatch_openjarvis_tool(tool, arguments or {}, executor_holder)
|
||||
|
||||
worker = to_worker_dict(tool)
|
||||
prompt = ""
|
||||
for key in ("input", "query", "code"):
|
||||
val = arguments.get(key)
|
||||
if isinstance(val, str) and val.strip():
|
||||
prompt = val
|
||||
break
|
||||
if not prompt.strip():
|
||||
# The orchestrator emitted a tool call with no/empty input. Don't hit
|
||||
# the API (it 400s on empty content) — return a usable error so the
|
||||
# rollout keeps going instead of dropping.
|
||||
return (f"[{tool.name}: no input provided — supply a non-empty "
|
||||
"'input' to delegate]", 0.0, 0, True)
|
||||
text, p, c, is_local, extra_cost, _n = _call_worker(worker, prompt, cfg)
|
||||
usd = (0.0 if is_local else _model_cost(str(tool.model), p, c)) + float(extra_cost)
|
||||
return text, usd, int(p) + int(c), bool(is_local)
|
||||
|
||||
def dispatch(tool: ExpertTool, arguments: Dict[str, Any]):
|
||||
# One span per route so the trace tree shows which expert/tool was called,
|
||||
# with the delegated input, the returned observation, and cost/tokens. The
|
||||
# wrapped OpenAI/Anthropic call (if any) nests one level deeper.
|
||||
# Span is titled with the REAL model that actually ran (``tool.model``) so
|
||||
# the trace reads clearly even under anonymization; ``anon_label`` in the
|
||||
# metadata records the opaque label the orchestrator actually saw/chose.
|
||||
real_model = str(tool.model)
|
||||
with span(f"route:{real_model}", span_type="tool", input=arguments,
|
||||
metadata={"real_model": real_model, "anon_label": tool.name,
|
||||
"backend": tool.backend_type}) as s:
|
||||
obs, usd, toks, is_local = _dispatch_inner(tool, arguments)
|
||||
s.log(output=obs, metrics={"cost_usd": usd, "tokens": toks},
|
||||
metadata={"is_local": is_local})
|
||||
return obs, usd, toks, is_local
|
||||
|
||||
return dispatch
|
||||
|
||||
|
||||
def teacher_rollout(
|
||||
question: str,
|
||||
tools: List[ExpertTool],
|
||||
*,
|
||||
teacher_model: str,
|
||||
base_url: Optional[str] = None,
|
||||
api_key: Optional[str] = None,
|
||||
temperature: float = 1.0,
|
||||
max_turns: int = 50,
|
||||
cfg: Optional[Dict[str, Any]] = None,
|
||||
) -> UnifiedRollout:
|
||||
"""Convenience: one full teacher rollout with real backends."""
|
||||
return run_unified_rollout(
|
||||
question,
|
||||
tools,
|
||||
call_orchestrator=make_call_orchestrator(
|
||||
teacher_model, base_url=base_url, api_key=api_key, temperature=temperature,
|
||||
),
|
||||
dispatch=make_dispatch(cfg),
|
||||
max_turns=max_turns,
|
||||
)
|
||||
|
||||
|
||||
__all__ = ["make_call_orchestrator", "make_dispatch", "teacher_rollout"]
|
||||
@@ -0,0 +1,422 @@
|
||||
"""Worker pool resolution + dispatch for ToolOrchestraAgent."""
|
||||
|
||||
from __future__ import annotations
|
||||
|
||||
import os
|
||||
from pathlib import Path
|
||||
from typing import Any, Dict, List, Optional, Tuple
|
||||
|
||||
from openjarvis.agents.hybrid._base import (
|
||||
ANTHROPIC_WEB_SEARCH_TOOL,
|
||||
WEB_SEARCH_COST_PER_CALL,
|
||||
LocalCloudAgent,
|
||||
)
|
||||
from openjarvis.agents.hybrid._prices import (
|
||||
PRICES,
|
||||
is_gpt5_family,
|
||||
supports_temperature,
|
||||
)
|
||||
from openjarvis.agents.hybrid.mini_swe_agent import run_swe_agent_loop
|
||||
from openjarvis.agents.hybrid.toolorchestra.experts import (
|
||||
_PAPER_CODER_OPENROUTER,
|
||||
_PAPER_GENERALIST_TIER3_OPENROUTER,
|
||||
)
|
||||
from openjarvis.agents.hybrid.toolorchestra.sandbox import (
|
||||
_call_modal_python,
|
||||
_call_tavily_search,
|
||||
)
|
||||
|
||||
def _paper_pool(
|
||||
local_model: Optional[str],
|
||||
local_endpoint: Optional[str],
|
||||
) -> List[Dict[str, Any]]:
|
||||
"""Paper-match worker pool (registered for traces / inspection).
|
||||
|
||||
NOTE: in RL mode the orchestrator dispatches via tool/slot rather than
|
||||
worker_id, so this list is purely informational — `_paper_expert_for`
|
||||
is the actual routing function. We still return a list here so the
|
||||
paradigm's trace metadata has something concrete to log.
|
||||
"""
|
||||
pool: List[Dict[str, Any]] = []
|
||||
if local_model and local_endpoint:
|
||||
pool.append({
|
||||
"id": len(pool),
|
||||
"name": "local-qwen",
|
||||
"type": "vllm",
|
||||
"model": local_model,
|
||||
"base_url": local_endpoint,
|
||||
"description": "Local Qwen vLLM (paper uses Qwen3-32B).",
|
||||
})
|
||||
pool.append({
|
||||
"id": len(pool), "name": "tavily-search",
|
||||
"type": "tavily-search", "model": "tavily",
|
||||
"description": "Tavily web search.",
|
||||
})
|
||||
pool.append({
|
||||
"id": len(pool), "name": "modal-python",
|
||||
"type": "modal-python", "model": "modal-python",
|
||||
"description": "Modal Sandbox for one-shot Python exec.",
|
||||
})
|
||||
pool.append({
|
||||
"id": len(pool), "name": "code-specialist",
|
||||
"type": "openrouter", "model": _PAPER_CODER_OPENROUTER,
|
||||
"description": "Qwen-2.5-Coder-32B via OpenRouter (paper).",
|
||||
})
|
||||
pool.append({
|
||||
"id": len(pool), "name": "generalist-llama",
|
||||
"type": "openrouter", "model": _PAPER_GENERALIST_TIER3_OPENROUTER,
|
||||
"description": "Llama-3.3-70B-Instruct via OpenRouter (paper tier-3).",
|
||||
})
|
||||
pool.append({
|
||||
"id": len(pool), "name": "generalist-gpt5",
|
||||
"type": "openai", "model": "gpt-5",
|
||||
"description": "GPT-5 frontier generalist.",
|
||||
})
|
||||
pool.append({
|
||||
"id": len(pool), "name": "generalist-gpt5-mini",
|
||||
"type": "openai", "model": "gpt-5-mini",
|
||||
"description": "GPT-5-mini mid generalist.",
|
||||
})
|
||||
return pool
|
||||
|
||||
def _default_pool(
|
||||
local_model: Optional[str],
|
||||
local_endpoint: Optional[str],
|
||||
cloud_model: str = "claude-opus-4-7",
|
||||
cloud_endpoint: str = "anthropic",
|
||||
) -> List[Dict[str, Any]]:
|
||||
"""Default heterogeneous worker pool.
|
||||
|
||||
The frontier worker's ``type`` + ``model`` track the cell's resolved
|
||||
``(cloud_model, cloud_endpoint)`` pair so non-Anthropic cells (gpt-5,
|
||||
gemini-2.5-pro, …) route their frontier slot to the right SDK.
|
||||
"""
|
||||
ep = (cloud_endpoint or "anthropic").lower()
|
||||
if ep not in ("anthropic", "openai", "gemini"):
|
||||
ep = "anthropic"
|
||||
pool: List[Dict[str, Any]] = []
|
||||
if local_model and local_endpoint:
|
||||
pool.append({
|
||||
"id": len(pool),
|
||||
"name": "local-qwen",
|
||||
"type": "vllm",
|
||||
"model": local_model,
|
||||
"base_url": local_endpoint,
|
||||
"description": (
|
||||
"Open-weights Qwen3.5 served locally. Cheap and fast. Good at "
|
||||
"concise extraction, formatting, arithmetic on given data."
|
||||
),
|
||||
})
|
||||
pool.append({
|
||||
"id": len(pool),
|
||||
"name": "web-search",
|
||||
"type": "anthropic-web-search",
|
||||
"model": "claude-haiku-4-5",
|
||||
"description": (
|
||||
"Anthropic server-side web_search. Use for facts that need a lookup "
|
||||
"(recent events, rare names/dates, niche sources). Returns a digest."
|
||||
),
|
||||
})
|
||||
pool.append({
|
||||
"id": len(pool),
|
||||
"name": f"frontier-{ep}",
|
||||
"type": ep,
|
||||
"model": cloud_model,
|
||||
"description": (
|
||||
"Frontier reasoning model. Use for hard multi-step reasoning, "
|
||||
"code review, or a final synthesis pass. Expensive — use sparingly."
|
||||
),
|
||||
})
|
||||
pool.append({
|
||||
"id": len(pool),
|
||||
"name": "frontier-openai-mini",
|
||||
"type": "openai",
|
||||
"model": "gpt-5-mini",
|
||||
"description": (
|
||||
"Mid-tier OpenAI model. Solid general knowledge and reasoning at a "
|
||||
"fraction of frontier cost."
|
||||
),
|
||||
})
|
||||
return pool
|
||||
|
||||
|
||||
# Worker types toolorchestra's `_call_worker` actually dispatches.
|
||||
#
|
||||
# Paper-match additions (2026-05-19) — opt in via `method_cfg.pool = "paper"`:
|
||||
# `tavily-search` — Tavily API search (the paper's web tool).
|
||||
# `openrouter` — OpenAI-compatible client at openrouter.ai/api/v1.
|
||||
# Used for the code/math specialists and Llama-3.3-70B /
|
||||
# Qwen3-32B generalists.
|
||||
# `modal-python` — One-shot Python exec in a fresh Modal Sandbox (the
|
||||
# paper's "Python sandbox" inside `enhance_reasoning`).
|
||||
_TOOLORCH_VALID_TYPES = (
|
||||
"vllm", "openai", "anthropic", "anthropic-web-search", "gemini",
|
||||
"tavily-search", "openrouter", "modal-python",
|
||||
)
|
||||
|
||||
# Default model used when an `anthropic-web-search` entry omits `model`.
|
||||
_DEFAULT_WEB_SEARCH_MODEL = "claude-haiku-4-5"
|
||||
|
||||
|
||||
def _resolve_worker_pool(
|
||||
cfg: Dict[str, Any],
|
||||
local_model: Optional[str],
|
||||
local_endpoint: Optional[str],
|
||||
cloud_model: str,
|
||||
cloud_endpoint: str = "anthropic",
|
||||
) -> List[Dict[str, Any]]:
|
||||
"""Return the worker pool for this run.
|
||||
|
||||
Strict replace, not merge: if ``cfg["worker_pool"]`` is set, the
|
||||
default pool is ignored entirely. Falls back to ``_default_pool`` when
|
||||
the override is absent.
|
||||
|
||||
Each user-supplied entry must be a dict with keys ``id``, ``name``,
|
||||
``type``, and (for non-search types) ``model``. ``type`` must be one
|
||||
of ``vllm`` / ``openai`` / ``anthropic`` / ``anthropic-web-search``.
|
||||
``anthropic-web-search`` entries may omit ``model`` — it defaults to
|
||||
``claude-haiku-4-5``.
|
||||
|
||||
Substitution: ``model = "$local"`` (or ``"<local>"``) resolves to
|
||||
``local_model``; ``model = "$cloud"`` / ``"<cloud>"`` to ``cloud_model``.
|
||||
|
||||
On any validation failure, raises ``ValueError`` with the message
|
||||
``"Invalid worker_pool entry [<id>]: <reason>"``. Fails fast at agent
|
||||
init rather than mid-task.
|
||||
"""
|
||||
override = cfg.get("worker_pool")
|
||||
if override is None:
|
||||
return _default_pool(local_model, local_endpoint, cloud_model, cloud_endpoint)
|
||||
if not isinstance(override, list) or not override:
|
||||
raise ValueError(
|
||||
"Invalid worker_pool entry [-]: worker_pool must be a non-empty list"
|
||||
)
|
||||
|
||||
resolved: List[Dict[str, Any]] = []
|
||||
seen_ids: set = set()
|
||||
has_non_search = False
|
||||
for raw in override:
|
||||
wid_repr = raw.get("id", "?") if isinstance(raw, dict) else "?"
|
||||
if not isinstance(raw, dict):
|
||||
raise ValueError(
|
||||
f"Invalid worker_pool entry [{wid_repr}]: entry must be a dict"
|
||||
)
|
||||
entry = dict(raw)
|
||||
wid = entry.get("id")
|
||||
if not isinstance(wid, int):
|
||||
raise ValueError(
|
||||
f"Invalid worker_pool entry [{wid_repr}]: 'id' must be an int"
|
||||
)
|
||||
if wid in seen_ids:
|
||||
raise ValueError(
|
||||
f"Invalid worker_pool entry [{wid}]: duplicate id"
|
||||
)
|
||||
seen_ids.add(wid)
|
||||
if not entry.get("name") or not isinstance(entry["name"], str):
|
||||
raise ValueError(
|
||||
f"Invalid worker_pool entry [{wid}]: 'name' must be a non-empty string"
|
||||
)
|
||||
wtype = entry.get("type") or entry.get("endpoint")
|
||||
if not isinstance(wtype, str) or wtype.lower() not in _TOOLORCH_VALID_TYPES:
|
||||
raise ValueError(
|
||||
f"Invalid worker_pool entry [{wid}]: 'type' must be one of "
|
||||
f"{_TOOLORCH_VALID_TYPES} (got {wtype!r})"
|
||||
)
|
||||
wtype = wtype.lower()
|
||||
entry["type"] = wtype
|
||||
# Substitute $local / $cloud placeholders (before any model check).
|
||||
model = entry.get("model")
|
||||
if isinstance(model, str) and model in ("$local", "<local>"):
|
||||
if not local_model:
|
||||
raise ValueError(
|
||||
f"Invalid worker_pool entry [{wid}]: model='{model}' "
|
||||
"requires a local_model to be configured for this cell"
|
||||
)
|
||||
model = local_model
|
||||
entry["model"] = model
|
||||
elif isinstance(model, str) and model in ("$cloud", "<cloud>"):
|
||||
model = cloud_model
|
||||
entry["model"] = model
|
||||
if wtype == "anthropic-web-search":
|
||||
if model in (None, ""):
|
||||
model = _DEFAULT_WEB_SEARCH_MODEL
|
||||
entry["model"] = model
|
||||
elif not isinstance(model, str):
|
||||
raise ValueError(
|
||||
f"Invalid worker_pool entry [{wid}]: 'model' must be a string when set"
|
||||
)
|
||||
# Search workers don't satisfy the "needs a solver" requirement.
|
||||
else:
|
||||
if not isinstance(model, str) or not model:
|
||||
raise ValueError(
|
||||
f"Invalid worker_pool entry [{wid}]: 'model' must be a non-empty string"
|
||||
)
|
||||
if wtype == "vllm":
|
||||
if not entry.get("base_url"):
|
||||
if not local_endpoint:
|
||||
raise ValueError(
|
||||
f"Invalid worker_pool entry [{wid}]: vllm worker needs "
|
||||
"'base_url' (or a configured local_endpoint to fall back to)"
|
||||
)
|
||||
entry["base_url"] = local_endpoint
|
||||
entry.setdefault("api_key", "EMPTY")
|
||||
else:
|
||||
if model not in PRICES:
|
||||
raise ValueError(
|
||||
f"Invalid worker_pool entry [{wid}]: model {model!r} "
|
||||
f"is not in PRICES (known: {sorted(PRICES)})"
|
||||
)
|
||||
has_non_search = True
|
||||
entry.setdefault(
|
||||
"description",
|
||||
f"User-supplied {wtype} worker ({model}).",
|
||||
)
|
||||
resolved.append(entry)
|
||||
|
||||
if not has_non_search:
|
||||
raise ValueError(
|
||||
"Invalid worker_pool entry [-]: worker_pool must contain at least "
|
||||
"one non-search worker (vllm / openai / anthropic)"
|
||||
)
|
||||
return resolved
|
||||
|
||||
|
||||
def _call_worker(
|
||||
worker: Dict[str, Any], prompt: str, cfg: Dict[str, Any]
|
||||
) -> Tuple[str, int, int, bool, float, int]:
|
||||
"""Returns (text, p_tok, c_tok, is_local, extra_cost, n_web_searches)."""
|
||||
wtype = worker.get("type", "openai")
|
||||
max_tok = int(cfg.get("worker_max_tokens") or os.environ.get("OJ_WORKER_MAX_TOKENS", "4096"))
|
||||
temp = float(cfg.get("worker_temperature", 0.2))
|
||||
|
||||
if wtype == "vllm":
|
||||
text, p, c = LocalCloudAgent._call_vllm(
|
||||
worker["model"],
|
||||
worker["base_url"],
|
||||
user=prompt,
|
||||
max_tokens=max_tok,
|
||||
temperature=temp,
|
||||
enable_thinking=False,
|
||||
)
|
||||
return text, p, c, True, 0.0, 0
|
||||
if wtype == "openai":
|
||||
is_gpt5 = is_gpt5_family(worker["model"])
|
||||
eff_temp = 1.0 if is_gpt5 else temp
|
||||
# GPT-5 is a reasoning model: hidden reasoning tokens count against
|
||||
# `max_completion_tokens`, so a 4096 cap can be fully consumed by
|
||||
# reasoning and leave 0 visible content (empty answer). Give the
|
||||
# reasoning headroom on top of the answer budget.
|
||||
eff_max_tok = max(max_tok, 16384) if is_gpt5 else max_tok
|
||||
text, p, c = LocalCloudAgent._call_openai(
|
||||
worker["model"],
|
||||
user=prompt,
|
||||
max_tokens=eff_max_tok,
|
||||
temperature=eff_temp,
|
||||
)
|
||||
return text, p, c, False, 0.0, 0
|
||||
if wtype == "gemini":
|
||||
text, p, c = LocalCloudAgent._call_gemini(
|
||||
worker["model"],
|
||||
user=prompt,
|
||||
max_tokens=max_tok,
|
||||
temperature=temp,
|
||||
)
|
||||
return text, p, c, False, 0.0, 0
|
||||
if wtype == "anthropic":
|
||||
eff_temp = temp if supports_temperature(worker["model"]) else 0.0
|
||||
text, p, c, _ = LocalCloudAgent._call_anthropic(
|
||||
worker["model"],
|
||||
user=prompt,
|
||||
max_tokens=max_tok,
|
||||
temperature=eff_temp,
|
||||
)
|
||||
return text, p, c, False, 0.0, 0
|
||||
if wtype == "anthropic-web-search":
|
||||
eff_temp = temp if supports_temperature(worker["model"]) else 0.0
|
||||
text, p, c, n_searches = LocalCloudAgent._call_anthropic(
|
||||
worker["model"],
|
||||
user=prompt,
|
||||
max_tokens=max_tok,
|
||||
temperature=eff_temp,
|
||||
tools=[ANTHROPIC_WEB_SEARCH_TOOL],
|
||||
tool_choice={"type": "any"},
|
||||
)
|
||||
extra = n_searches * WEB_SEARCH_COST_PER_CALL
|
||||
return text, p, c, False, extra, n_searches
|
||||
if wtype == "tavily-search":
|
||||
# Tavily costs are flat per call; charge `WEB_SEARCH_COST_PER_CALL`
|
||||
# for parity with the Anthropic web-search worker. One call = one
|
||||
# "n_search" for accounting.
|
||||
max_results = int(cfg.get("tavily_max_results", 5))
|
||||
text, p, c = _call_tavily_search(str(prompt), max_results=max_results)
|
||||
return text, p, c, False, WEB_SEARCH_COST_PER_CALL, 1
|
||||
if wtype == "openrouter":
|
||||
text, p, c = LocalCloudAgent._call_openrouter(
|
||||
worker["model"],
|
||||
user=prompt,
|
||||
max_tokens=max_tok,
|
||||
temperature=temp,
|
||||
)
|
||||
return text, p, c, False, 0.0, 0
|
||||
if wtype == "modal-python":
|
||||
# `prompt` is the python code string to exec.
|
||||
timeout_s = int(cfg.get("modal_python_timeout_s", 60))
|
||||
out, _rc = _call_modal_python(str(prompt), timeout_s=timeout_s)
|
||||
# No LLM tokens consumed; report 0 in/out. Cost is whatever Modal
|
||||
# charges per sandbox-second — not tracked here.
|
||||
return out, 0, 0, False, 0.0, 0
|
||||
raise ValueError(f"unsupported worker type: {wtype!r}")
|
||||
|
||||
|
||||
def _swe_call_worker(
|
||||
worker: Dict[str, Any],
|
||||
prompt: str,
|
||||
cfg: Dict[str, Any],
|
||||
task: Dict[str, Any],
|
||||
workdir: Path,
|
||||
turn: int,
|
||||
) -> Tuple[str, int, int, bool, float, int, int]:
|
||||
"""SWE-bench worker dispatch: route solver workers through
|
||||
run_swe_agent_loop on a shared workdir. Web-search workers fall back
|
||||
to the regular one-shot dispatch (search isn't an agent loop).
|
||||
|
||||
Trailing ``bash_turns`` (last element) counts agent-loop turns so the
|
||||
caller can surface ``tool_calls`` per row. Fallbacks to one-shot
|
||||
workers return 0 bash turns (no agent loop ran)."""
|
||||
wtype = worker.get("type", "openai")
|
||||
if wtype == "anthropic-web-search":
|
||||
# Search workers stay one-shot.
|
||||
text, p, c, is_local, extra, n_searches = _call_worker(worker, prompt, cfg)
|
||||
return text, p, c, is_local, extra, n_searches, 0
|
||||
if wtype == "vllm":
|
||||
backbone = "local"
|
||||
endpoint = worker.get("base_url")
|
||||
loop_cloud_endpoint = "anthropic" # unused when backbone=local
|
||||
elif wtype in ("anthropic", "openai", "gemini"):
|
||||
backbone = "cloud"
|
||||
endpoint = None
|
||||
loop_cloud_endpoint = wtype
|
||||
else:
|
||||
# Unknown type — one-shot fallback.
|
||||
text, p, c, is_local, extra, n_searches = _call_worker(worker, prompt, cfg)
|
||||
return text, p, c, is_local, extra, n_searches, 0
|
||||
out = run_swe_agent_loop(
|
||||
task,
|
||||
backbone=backbone,
|
||||
backbone_model=worker["model"],
|
||||
cloud_endpoint=loop_cloud_endpoint,
|
||||
local_endpoint=endpoint,
|
||||
initial_prompt=prompt,
|
||||
max_turns=int(cfg.get("swe_max_turns", 30)),
|
||||
bash_timeout=int(cfg.get("swe_bash_timeout_s", 120)),
|
||||
output_cap=int(cfg.get("swe_output_cap", 10_000)),
|
||||
turn_max_tokens=int(cfg.get("swe_turn_max_tokens", 4096)),
|
||||
trace_prefix=f"toolorch_turn{turn}",
|
||||
workdir=workdir,
|
||||
)
|
||||
is_local = backbone == "local"
|
||||
return (
|
||||
out["final_summary"] or out["answer"],
|
||||
out["tokens_in"], out["tokens_out"],
|
||||
is_local, 0.0, 0, int(out["turns"]),
|
||||
)
|
||||
@@ -13,6 +13,7 @@ Supports two modes:
|
||||
from __future__ import annotations
|
||||
|
||||
import concurrent.futures
|
||||
import json
|
||||
import re
|
||||
from typing import Any, List, Optional
|
||||
|
||||
@@ -201,6 +202,70 @@ class OrchestratorAgent(ToolUsingAgent):
|
||||
|
||||
return result
|
||||
|
||||
@staticmethod
|
||||
def _boxed_tool_call(content, openai_tools, call_id):
|
||||
r"""Salvage a tool call the model wrote as ``\boxed{tool: query}``.
|
||||
|
||||
Some SFT'd checkpoints (Qwen reasoning-mode habit) emit their action
|
||||
as a boxed string instead of a real ``<tool_call>`` — e.g.
|
||||
``\boxed{web_search: latest news}`` or ``\boxed{web_search query: ...}``.
|
||||
The engine's tool-call parser sees no structured call, so without this
|
||||
the agent would grade the half-finished thought as its final answer and
|
||||
never actually act. Here we recover ``(tool, args)`` and build a real
|
||||
``ToolCall`` keyed on the tool's primary parameter. Returns ``None`` when
|
||||
the boxed content isn't a recognisable tool action (leaving genuine
|
||||
boxed final answers untouched).
|
||||
"""
|
||||
if not openai_tools or "\\boxed" not in content:
|
||||
return None
|
||||
|
||||
# tool name -> primary parameter (first required, else first property)
|
||||
param_of: dict[str, str] = {}
|
||||
for t in openai_tools:
|
||||
fn = t.get("function", t)
|
||||
name = fn.get("name")
|
||||
if not name:
|
||||
continue
|
||||
params = fn.get("parameters") or {}
|
||||
props = list((params.get("properties") or {}).keys())
|
||||
req = params.get("required") or []
|
||||
param_of[name] = req[0] if req else (props[0] if props else "query")
|
||||
|
||||
# The action is the LAST boxed expression in the response.
|
||||
boxed = re.findall(r"\\boxed\{([^{}]*)\}", content)
|
||||
if not boxed:
|
||||
return None
|
||||
inner = boxed[-1].strip()
|
||||
|
||||
# Match a leading tool name (longest first to avoid prefix collisions),
|
||||
# then strip the separator between name and args. Handles
|
||||
# ``tool: q`` / ``tool query: q`` / ``tool q`` / ``tool(q)``.
|
||||
for name in sorted(param_of, key=len, reverse=True):
|
||||
m = re.match(
|
||||
r"^\s*" + re.escape(name) + r"\b(.*)$",
|
||||
inner,
|
||||
re.DOTALL | re.IGNORECASE,
|
||||
)
|
||||
if not m:
|
||||
continue
|
||||
rest = m.group(1).strip()
|
||||
if rest.startswith("(") and rest.endswith(")"):
|
||||
rest = rest[1:-1] # tool(args) wrapper -> drop matching parens
|
||||
else:
|
||||
# optional "query" word, then a ":" / "-" separator
|
||||
rest = re.sub(
|
||||
r"^\s*(?:query\s*)?[:\-]\s*", "", rest, flags=re.IGNORECASE
|
||||
)
|
||||
arg = rest.strip().strip("\"'").strip()
|
||||
if not arg:
|
||||
return None
|
||||
return ToolCall(
|
||||
id=call_id,
|
||||
name=name,
|
||||
arguments=json.dumps({param_of[name]: arg}),
|
||||
)
|
||||
return None
|
||||
|
||||
# ------------------------------------------------------------------
|
||||
# Function-calling mode (original behaviour)
|
||||
# ------------------------------------------------------------------
|
||||
@@ -245,8 +310,31 @@ class OrchestratorAgent(ToolUsingAgent):
|
||||
content = result.get("content", "")
|
||||
raw_tool_calls = result.get("tool_calls", [])
|
||||
|
||||
# No tool calls -> check continuation, then final answer
|
||||
# No tool calls -> try to salvage a \boxed{tool: query} action the
|
||||
# model wrote instead of a real <tool_call>; else final answer.
|
||||
if not raw_tool_calls:
|
||||
boxed_call = self._boxed_tool_call(
|
||||
content, openai_tools, f"boxed_{turns}"
|
||||
)
|
||||
if boxed_call is not None:
|
||||
messages.append(
|
||||
Message(
|
||||
role=Role.ASSISTANT,
|
||||
content=content,
|
||||
tool_calls=[boxed_call],
|
||||
)
|
||||
)
|
||||
tool_result = self._executor.execute(boxed_call)
|
||||
all_tool_results.append(tool_result)
|
||||
messages.append(
|
||||
Message(
|
||||
role=Role.TOOL,
|
||||
content=tool_result.content,
|
||||
tool_call_id=boxed_call.id,
|
||||
name=boxed_call.name,
|
||||
)
|
||||
)
|
||||
continue
|
||||
content = self._check_continuation(result, messages)
|
||||
content = self._strip_think_tags(content)
|
||||
self._emit_turn_end(turns=turns, content_length=len(content))
|
||||
|
||||
@@ -1408,7 +1408,7 @@ class OptimizeConfig:
|
||||
optimizer_provider: str = "anthropic"
|
||||
benchmark: str = ""
|
||||
max_samples: int = 50
|
||||
judge_model: str = "gpt-5-mini-2025-08-07"
|
||||
judge_model: str = "claude-haiku-4-5-20251001"
|
||||
db_path: str = str(DEFAULT_CONFIG_DIR / "optimize.db")
|
||||
|
||||
|
||||
|
||||
@@ -1141,7 +1141,7 @@ def main():
|
||||
"-n", "--max-samples", type=int, default=None, help="Maximum samples to evaluate"
|
||||
)
|
||||
@click.option("-w", "--max-workers", type=int, default=4, help="Parallel workers")
|
||||
@click.option("--judge-model", default="gpt-5-mini-2025-08-07", help="LLM judge model")
|
||||
@click.option("--judge-model", default="claude-haiku-4-5-20251001", help="LLM judge model")
|
||||
@click.option("-o", "--output", "output_path", default=None, help="Output JSONL path")
|
||||
@click.option("--seed", type=int, default=42, help="Random seed")
|
||||
@click.option("--split", "dataset_split", default=None, help="Dataset split override")
|
||||
@@ -1387,7 +1387,7 @@ def run(
|
||||
"-n", "--max-samples", type=int, default=None, help="Max samples per benchmark"
|
||||
)
|
||||
@click.option("-w", "--max-workers", type=int, default=4, help="Parallel workers")
|
||||
@click.option("--judge-model", default="gpt-5-mini-2025-08-07", help="LLM judge model")
|
||||
@click.option("--judge-model", default="claude-haiku-4-5-20251001", help="LLM judge model")
|
||||
@click.option("--output-dir", default="results/", help="Output directory for results")
|
||||
@click.option("--seed", type=int, default=42, help="Random seed")
|
||||
@click.option("-v", "--verbose", is_flag=True, help="Verbose logging")
|
||||
|
||||
@@ -98,7 +98,7 @@ BENCHMARKS: Dict[str, Dict[str, object]] = {
|
||||
},
|
||||
}
|
||||
|
||||
DEFAULT_JUDGE_MODEL = "gpt-5-mini-2025-08-07"
|
||||
DEFAULT_JUDGE_MODEL = "claude-haiku-4-5-20251001"
|
||||
|
||||
|
||||
def _template_path() -> Path:
|
||||
|
||||
@@ -11,7 +11,7 @@ temperature = 0.6
|
||||
max_tokens = 32768
|
||||
|
||||
[judge]
|
||||
model = "gpt-5-mini-2025-08-07"
|
||||
model = "claude-haiku-4-5-20251001"
|
||||
temperature = 0.0
|
||||
max_tokens = 1024
|
||||
|
||||
|
||||
@@ -8,7 +8,7 @@ temperature = 0.6
|
||||
max_tokens = 8192
|
||||
|
||||
[judge]
|
||||
model = "gpt-5-mini-2025-08-07"
|
||||
model = "claude-haiku-4-5-20251001"
|
||||
temperature = 0.0
|
||||
max_tokens = 4096
|
||||
|
||||
|
||||
@@ -8,7 +8,7 @@ temperature = 0.6
|
||||
max_tokens = 8192
|
||||
|
||||
[judge]
|
||||
model = "gpt-5-mini-2025-08-07"
|
||||
model = "claude-haiku-4-5-20251001"
|
||||
temperature = 0.0
|
||||
max_tokens = 4096
|
||||
engine = "cloud"
|
||||
|
||||
@@ -8,7 +8,7 @@ temperature = 0.6
|
||||
max_tokens = 8192
|
||||
|
||||
[judge]
|
||||
model = "gpt-5-mini-2025-08-07"
|
||||
model = "claude-haiku-4-5-20251001"
|
||||
temperature = 0.0
|
||||
max_tokens = 4096
|
||||
|
||||
|
||||
@@ -8,7 +8,7 @@ temperature = 0.6
|
||||
max_tokens = 16384
|
||||
|
||||
[judge]
|
||||
model = "gpt-5-mini-2025-08-07"
|
||||
model = "claude-haiku-4-5-20251001"
|
||||
temperature = 0.0
|
||||
max_tokens = 4096
|
||||
|
||||
|
||||
@@ -8,7 +8,7 @@ temperature = 0.6
|
||||
max_tokens = 8192
|
||||
|
||||
[judge]
|
||||
model = "gpt-5-mini-2025-08-07"
|
||||
model = "claude-haiku-4-5-20251001"
|
||||
temperature = 0.0
|
||||
engine = "cloud"
|
||||
|
||||
|
||||
@@ -8,7 +8,7 @@ temperature = 1.0
|
||||
max_tokens = 8192
|
||||
|
||||
[judge]
|
||||
model = "gpt-5-mini-2025-08-07"
|
||||
model = "claude-haiku-4-5-20251001"
|
||||
temperature = 0.0
|
||||
max_tokens = 4096
|
||||
engine = "cloud"
|
||||
|
||||
@@ -8,7 +8,7 @@ temperature = 0.6
|
||||
max_tokens = 8192
|
||||
|
||||
[judge]
|
||||
model = "gpt-5-mini-2025-08-07"
|
||||
model = "claude-haiku-4-5-20251001"
|
||||
temperature = 0.0
|
||||
engine = "cloud"
|
||||
|
||||
|
||||
@@ -8,7 +8,7 @@ temperature = 0.6
|
||||
max_tokens = 16384
|
||||
|
||||
[judge]
|
||||
model = "gpt-5-mini-2025-08-07"
|
||||
model = "claude-haiku-4-5-20251001"
|
||||
temperature = 0.0
|
||||
max_tokens = 4096
|
||||
|
||||
|
||||
@@ -8,7 +8,7 @@ temperature = 0.6
|
||||
max_tokens = 8192
|
||||
|
||||
[judge]
|
||||
model = "gpt-5-mini-2025-08-07"
|
||||
model = "claude-haiku-4-5-20251001"
|
||||
temperature = 0.0
|
||||
max_tokens = 4096
|
||||
engine = "cloud"
|
||||
|
||||
@@ -8,7 +8,7 @@ temperature = 0.6
|
||||
max_tokens = 8192
|
||||
|
||||
[judge]
|
||||
model = "gpt-5-mini-2025-08-07"
|
||||
model = "claude-haiku-4-5-20251001"
|
||||
temperature = 0.0
|
||||
engine = "cloud"
|
||||
|
||||
|
||||
@@ -8,7 +8,7 @@ temperature = 0.6
|
||||
max_tokens = 8192
|
||||
|
||||
[judge]
|
||||
model = "gpt-5-mini-2025-08-07"
|
||||
model = "claude-haiku-4-5-20251001"
|
||||
temperature = 0.0
|
||||
max_tokens = 4096
|
||||
engine = "cloud"
|
||||
|
||||
@@ -8,7 +8,7 @@ temperature = 0.6
|
||||
max_tokens = 8192
|
||||
|
||||
[judge]
|
||||
model = "gpt-5-mini-2025-08-07"
|
||||
model = "claude-haiku-4-5-20251001"
|
||||
temperature = 0.0
|
||||
max_tokens = 4096
|
||||
|
||||
|
||||
@@ -8,7 +8,7 @@ temperature = 0.6
|
||||
max_tokens = 8192
|
||||
|
||||
[judge]
|
||||
model = "gpt-5-mini-2025-08-07"
|
||||
model = "claude-haiku-4-5-20251001"
|
||||
temperature = 0.0
|
||||
max_tokens = 4096
|
||||
|
||||
|
||||
@@ -8,7 +8,7 @@ temperature = 0.6
|
||||
max_tokens = 8192
|
||||
|
||||
[judge]
|
||||
model = "gpt-5-mini-2025-08-07"
|
||||
model = "claude-haiku-4-5-20251001"
|
||||
temperature = 0.0
|
||||
max_tokens = 4096
|
||||
|
||||
|
||||
@@ -8,7 +8,7 @@ temperature = 0.6
|
||||
max_tokens = 8192
|
||||
|
||||
[judge]
|
||||
model = "gpt-5-mini-2025-08-07"
|
||||
model = "claude-haiku-4-5-20251001"
|
||||
temperature = 0.0
|
||||
max_tokens = 4096
|
||||
|
||||
|
||||
@@ -8,7 +8,7 @@ temperature = 0.6
|
||||
max_tokens = 8192
|
||||
|
||||
[judge]
|
||||
model = "gpt-5-mini-2025-08-07"
|
||||
model = "claude-haiku-4-5-20251001"
|
||||
temperature = 0.0
|
||||
max_tokens = 4096
|
||||
|
||||
|
||||
@@ -8,7 +8,7 @@ temperature = 0.6
|
||||
max_tokens = 8192
|
||||
|
||||
[judge]
|
||||
model = "gpt-5-mini-2025-08-07"
|
||||
model = "claude-haiku-4-5-20251001"
|
||||
temperature = 0.0
|
||||
max_tokens = 4096
|
||||
|
||||
|
||||
@@ -8,7 +8,7 @@ temperature = 0.6
|
||||
max_tokens = 8192
|
||||
|
||||
[judge]
|
||||
model = "gpt-5-mini-2025-08-07"
|
||||
model = "claude-haiku-4-5-20251001"
|
||||
temperature = 0.0
|
||||
max_tokens = 4096
|
||||
|
||||
|
||||
@@ -8,7 +8,7 @@ temperature = 0.6
|
||||
max_tokens = 8192
|
||||
|
||||
[judge]
|
||||
model = "gpt-5-mini-2025-08-07"
|
||||
model = "claude-haiku-4-5-20251001"
|
||||
temperature = 0.0
|
||||
max_tokens = 4096
|
||||
|
||||
|
||||
@@ -8,7 +8,7 @@ temperature = 0.6
|
||||
max_tokens = 8192
|
||||
|
||||
[judge]
|
||||
model = "gpt-5-mini-2025-08-07"
|
||||
model = "claude-haiku-4-5-20251001"
|
||||
temperature = 0.0
|
||||
max_tokens = 4096
|
||||
|
||||
|
||||
@@ -8,7 +8,7 @@ temperature = 0.6
|
||||
max_tokens = 8192
|
||||
|
||||
[judge]
|
||||
model = "gpt-5-mini-2025-08-07"
|
||||
model = "claude-haiku-4-5-20251001"
|
||||
temperature = 0.0
|
||||
max_tokens = 4096
|
||||
|
||||
|
||||
@@ -8,7 +8,7 @@ temperature = 0.6
|
||||
max_tokens = 16384
|
||||
|
||||
[judge]
|
||||
model = "gpt-5-mini-2025-08-07"
|
||||
model = "claude-haiku-4-5-20251001"
|
||||
temperature = 0.0
|
||||
max_tokens = 4096
|
||||
|
||||
|
||||
@@ -10,7 +10,7 @@ temperature = 0.0
|
||||
max_tokens = 2048
|
||||
|
||||
[judge]
|
||||
model = "gpt-5-mini-2025-08-07"
|
||||
model = "claude-haiku-4-5-20251001"
|
||||
temperature = 0.0
|
||||
max_tokens = 2048
|
||||
|
||||
|
||||
@@ -10,7 +10,7 @@ temperature = 0.0
|
||||
max_tokens = 2048
|
||||
|
||||
[judge]
|
||||
model = "gpt-5-mini-2025-08-07"
|
||||
model = "claude-haiku-4-5-20251001"
|
||||
temperature = 0.0
|
||||
max_tokens = 1024
|
||||
|
||||
|
||||
@@ -10,7 +10,7 @@ temperature = 0.0
|
||||
max_tokens = 2048
|
||||
|
||||
[judge]
|
||||
model = "gpt-5-mini-2025-08-07"
|
||||
model = "claude-haiku-4-5-20251001"
|
||||
temperature = 0.0
|
||||
max_tokens = 1024
|
||||
|
||||
|
||||
@@ -12,7 +12,7 @@ temperature = 0.6
|
||||
max_tokens = 32768
|
||||
|
||||
[judge]
|
||||
model = "gpt-5-mini-2025-08-07"
|
||||
model = "claude-haiku-4-5-20251001"
|
||||
temperature = 0.0
|
||||
max_tokens = 4096
|
||||
|
||||
|
||||
@@ -12,7 +12,7 @@ temperature = 1.0 # Kimi-K2.5 thinking mode recommended
|
||||
max_tokens = 32768
|
||||
|
||||
[judge]
|
||||
model = "gpt-5-mini-2025-08-07"
|
||||
model = "claude-haiku-4-5-20251001"
|
||||
temperature = 0.0
|
||||
max_tokens = 4096
|
||||
|
||||
|
||||
@@ -11,7 +11,7 @@ temperature = 0.0
|
||||
max_tokens = 4096
|
||||
|
||||
[judge]
|
||||
model = "gpt-5-mini-2025-08-07"
|
||||
model = "claude-haiku-4-5-20251001"
|
||||
temperature = 0.0
|
||||
max_tokens = 1024
|
||||
|
||||
|
||||
@@ -10,7 +10,7 @@ temperature = 0.0
|
||||
max_tokens = 4096
|
||||
|
||||
[judge]
|
||||
model = "gpt-5-mini-2025-08-07"
|
||||
model = "claude-haiku-4-5-20251001"
|
||||
temperature = 0.0
|
||||
engine = "cloud"
|
||||
|
||||
|
||||
@@ -10,7 +10,7 @@ temperature = 0.0
|
||||
max_tokens = 4096
|
||||
|
||||
[judge]
|
||||
model = "gpt-5-mini-2025-08-07"
|
||||
model = "claude-haiku-4-5-20251001"
|
||||
temperature = 0.0
|
||||
engine = "cloud"
|
||||
|
||||
|
||||
@@ -17,7 +17,7 @@ temperature = 0.0
|
||||
max_tokens = 4096
|
||||
|
||||
[judge]
|
||||
model = "gpt-5-mini-2025-08-07"
|
||||
model = "claude-haiku-4-5-20251001"
|
||||
temperature = 0.0
|
||||
engine = "cloud"
|
||||
|
||||
|
||||
@@ -14,7 +14,7 @@ temperature = 0.0
|
||||
max_tokens = 4096
|
||||
|
||||
[judge]
|
||||
model = "gpt-5-mini-2025-08-07"
|
||||
model = "claude-haiku-4-5-20251001"
|
||||
temperature = 0.0
|
||||
engine = "cloud"
|
||||
|
||||
|
||||
@@ -10,7 +10,7 @@ temperature = 0.0
|
||||
max_tokens = 4096
|
||||
|
||||
[judge]
|
||||
model = "gpt-5-mini-2025-08-07"
|
||||
model = "claude-haiku-4-5-20251001"
|
||||
temperature = 0.0
|
||||
engine = "cloud"
|
||||
|
||||
|
||||
@@ -8,7 +8,7 @@ temperature = 0.0
|
||||
max_tokens = 4096
|
||||
|
||||
[judge]
|
||||
model = "gpt-5-mini-2025-08-07"
|
||||
model = "claude-haiku-4-5-20251001"
|
||||
temperature = 0.0
|
||||
engine = "cloud"
|
||||
|
||||
|
||||
@@ -8,7 +8,7 @@ temperature = 0.0
|
||||
max_tokens = 4096
|
||||
|
||||
[judge]
|
||||
model = "gpt-5-mini-2025-08-07"
|
||||
model = "claude-haiku-4-5-20251001"
|
||||
temperature = 0.0
|
||||
engine = "cloud"
|
||||
|
||||
|
||||
@@ -10,7 +10,7 @@ temperature = 0.0
|
||||
max_tokens = 4096
|
||||
|
||||
[judge]
|
||||
model = "gpt-5-mini-2025-08-07"
|
||||
model = "claude-haiku-4-5-20251001"
|
||||
temperature = 0.0
|
||||
engine = "cloud"
|
||||
|
||||
|
||||
@@ -10,7 +10,7 @@ temperature = 0.0
|
||||
max_tokens = 4096
|
||||
|
||||
[judge]
|
||||
model = "gpt-5-mini-2025-08-07"
|
||||
model = "claude-haiku-4-5-20251001"
|
||||
temperature = 0.0
|
||||
engine = "cloud"
|
||||
|
||||
|
||||
@@ -8,7 +8,7 @@ temperature = 0.0
|
||||
max_tokens = 4096
|
||||
|
||||
[judge]
|
||||
model = "gpt-5-mini-2025-08-07"
|
||||
model = "claude-haiku-4-5-20251001"
|
||||
temperature = 0.0
|
||||
engine = "cloud"
|
||||
|
||||
|
||||
@@ -10,7 +10,7 @@ temperature = 0.0
|
||||
max_tokens = 4096
|
||||
|
||||
[judge]
|
||||
model = "gpt-5-mini-2025-08-07"
|
||||
model = "claude-haiku-4-5-20251001"
|
||||
temperature = 0.0
|
||||
engine = "cloud"
|
||||
|
||||
|
||||
@@ -8,7 +8,7 @@ temperature = 0.6
|
||||
max_tokens = 16384
|
||||
|
||||
[judge]
|
||||
model = "gpt-5-mini-2025-08-07"
|
||||
model = "claude-haiku-4-5-20251001"
|
||||
temperature = 0.0
|
||||
max_tokens = 4096
|
||||
|
||||
|
||||
@@ -8,7 +8,7 @@ temperature = 0.6
|
||||
max_tokens = 16384
|
||||
|
||||
[judge]
|
||||
model = "gpt-5-mini-2025-08-07"
|
||||
model = "claude-haiku-4-5-20251001"
|
||||
temperature = 0.0
|
||||
engine = "cloud"
|
||||
|
||||
|
||||
@@ -8,7 +8,7 @@ temperature = 0.6
|
||||
max_tokens = 16384
|
||||
|
||||
[judge]
|
||||
model = "gpt-5-mini-2025-08-07"
|
||||
model = "claude-haiku-4-5-20251001"
|
||||
temperature = 0.0
|
||||
max_tokens = 4096
|
||||
engine = "cloud"
|
||||
|
||||
@@ -8,7 +8,7 @@ temperature = 0.6
|
||||
max_tokens = 16384
|
||||
|
||||
[judge]
|
||||
model = "gpt-5-mini-2025-08-07"
|
||||
model = "claude-haiku-4-5-20251001"
|
||||
temperature = 0.0
|
||||
engine = "cloud"
|
||||
|
||||
|
||||
@@ -8,7 +8,7 @@ temperature = 0.6
|
||||
max_tokens = 16384
|
||||
|
||||
[judge]
|
||||
model = "gpt-5-mini-2025-08-07"
|
||||
model = "claude-haiku-4-5-20251001"
|
||||
temperature = 0.0
|
||||
engine = "cloud"
|
||||
|
||||
|
||||
@@ -8,7 +8,7 @@ temperature = 0.6
|
||||
max_tokens = 16384
|
||||
|
||||
[judge]
|
||||
model = "gpt-5-mini-2025-08-07"
|
||||
model = "claude-haiku-4-5-20251001"
|
||||
temperature = 0.0
|
||||
engine = "cloud"
|
||||
|
||||
|
||||
@@ -8,7 +8,7 @@ temperature = 0.6
|
||||
max_tokens = 16384
|
||||
|
||||
[judge]
|
||||
model = "gpt-5-mini-2025-08-07"
|
||||
model = "claude-haiku-4-5-20251001"
|
||||
temperature = 0.0
|
||||
engine = "cloud"
|
||||
|
||||
|
||||
@@ -8,7 +8,7 @@ temperature = 0.6
|
||||
max_tokens = 16384
|
||||
|
||||
[judge]
|
||||
model = "gpt-5-mini-2025-08-07"
|
||||
model = "claude-haiku-4-5-20251001"
|
||||
temperature = 0.0
|
||||
engine = "cloud"
|
||||
|
||||
|
||||
@@ -8,7 +8,7 @@ temperature = 0.6
|
||||
max_tokens = 16384
|
||||
|
||||
[judge]
|
||||
model = "gpt-5-mini-2025-08-07"
|
||||
model = "claude-haiku-4-5-20251001"
|
||||
temperature = 0.0
|
||||
engine = "cloud"
|
||||
|
||||
|
||||
@@ -8,7 +8,7 @@ temperature = 0.6
|
||||
max_tokens = 16384
|
||||
|
||||
[judge]
|
||||
model = "gpt-5-mini-2025-08-07"
|
||||
model = "claude-haiku-4-5-20251001"
|
||||
temperature = 0.0
|
||||
max_tokens = 4096
|
||||
|
||||
|
||||
@@ -8,7 +8,7 @@ temperature = 0.6
|
||||
max_tokens = 16384
|
||||
|
||||
[judge]
|
||||
model = "gpt-5-mini-2025-08-07"
|
||||
model = "claude-haiku-4-5-20251001"
|
||||
temperature = 0.0
|
||||
engine = "cloud"
|
||||
|
||||
|
||||
@@ -8,7 +8,7 @@ temperature = 0.6
|
||||
max_tokens = 16384
|
||||
|
||||
[judge]
|
||||
model = "gpt-5-mini-2025-08-07"
|
||||
model = "claude-haiku-4-5-20251001"
|
||||
temperature = 0.0
|
||||
engine = "cloud"
|
||||
|
||||
|
||||
@@ -8,7 +8,7 @@ temperature = 0.6
|
||||
max_tokens = 16384
|
||||
|
||||
[judge]
|
||||
model = "gpt-5-mini-2025-08-07"
|
||||
model = "claude-haiku-4-5-20251001"
|
||||
temperature = 0.0
|
||||
max_tokens = 4096
|
||||
|
||||
|
||||
@@ -8,7 +8,7 @@ temperature = 0.6
|
||||
max_tokens = 16384
|
||||
|
||||
[judge]
|
||||
model = "gpt-5-mini-2025-08-07"
|
||||
model = "claude-haiku-4-5-20251001"
|
||||
temperature = 0.0
|
||||
max_tokens = 4096
|
||||
|
||||
|
||||
@@ -8,7 +8,7 @@ temperature = 0.6
|
||||
max_tokens = 16384
|
||||
|
||||
[judge]
|
||||
model = "gpt-5-mini-2025-08-07"
|
||||
model = "claude-haiku-4-5-20251001"
|
||||
temperature = 0.0
|
||||
max_tokens = 4096
|
||||
engine = "cloud"
|
||||
|
||||
@@ -8,7 +8,7 @@ temperature = 0.6
|
||||
max_tokens = 16384
|
||||
|
||||
[judge]
|
||||
model = "gpt-5-mini-2025-08-07"
|
||||
model = "claude-haiku-4-5-20251001"
|
||||
temperature = 0.0
|
||||
max_tokens = 4096
|
||||
engine = "cloud"
|
||||
|
||||
@@ -8,7 +8,7 @@ temperature = 0.6
|
||||
max_tokens = 16384
|
||||
|
||||
[judge]
|
||||
model = "gpt-5-mini-2025-08-07"
|
||||
model = "claude-haiku-4-5-20251001"
|
||||
temperature = 0.0
|
||||
max_tokens = 4096
|
||||
engine = "cloud"
|
||||
|
||||
@@ -8,7 +8,7 @@ temperature = 0.6
|
||||
max_tokens = 16384
|
||||
|
||||
[judge]
|
||||
model = "gpt-5-mini-2025-08-07"
|
||||
model = "claude-haiku-4-5-20251001"
|
||||
temperature = 0.0
|
||||
max_tokens = 4096
|
||||
engine = "cloud"
|
||||
|
||||
@@ -8,7 +8,7 @@ temperature = 0.6
|
||||
max_tokens = 16384
|
||||
|
||||
[judge]
|
||||
model = "gpt-5-mini-2025-08-07"
|
||||
model = "claude-haiku-4-5-20251001"
|
||||
temperature = 0.0
|
||||
max_tokens = 4096
|
||||
engine = "cloud"
|
||||
|
||||
@@ -8,7 +8,7 @@ temperature = 0.6
|
||||
max_tokens = 16384
|
||||
|
||||
[judge]
|
||||
model = "gpt-5-mini-2025-08-07"
|
||||
model = "claude-haiku-4-5-20251001"
|
||||
temperature = 0.0
|
||||
engine = "cloud"
|
||||
|
||||
|
||||
@@ -8,7 +8,7 @@ temperature = 0.6
|
||||
max_tokens = 16384
|
||||
|
||||
[judge]
|
||||
model = "gpt-5-mini-2025-08-07"
|
||||
model = "claude-haiku-4-5-20251001"
|
||||
temperature = 0.0
|
||||
max_tokens = 4096
|
||||
engine = "cloud"
|
||||
|
||||
@@ -8,7 +8,7 @@ temperature = 0.6
|
||||
max_tokens = 16384
|
||||
|
||||
[judge]
|
||||
model = "gpt-5-mini-2025-08-07"
|
||||
model = "claude-haiku-4-5-20251001"
|
||||
temperature = 0.0
|
||||
max_tokens = 4096
|
||||
engine = "cloud"
|
||||
|
||||
@@ -8,7 +8,7 @@ temperature = 0.6
|
||||
max_tokens = 16384
|
||||
|
||||
[judge]
|
||||
model = "gpt-5-mini-2025-08-07"
|
||||
model = "claude-haiku-4-5-20251001"
|
||||
temperature = 0.0
|
||||
max_tokens = 4096
|
||||
engine = "cloud"
|
||||
|
||||
Some files were not shown because too many files have changed in this diff Show More
Reference in New Issue
Block a user