ResearchOS / scripts /critic_supervise.py
Kagzy's picture
v2 backend: figures, math prompts, live OpenAlex, 275k corpus via private dataset
e5c059f verified
Raw History Blame Contribute Delete
4.89 kB
"""Stage 2: build reference critiques for the trajectories (Algorithm 2).
python scripts/critic_supervise.py --limit 300 -k 3
Reference verdicts come from HotpotQA gold answers, so this script only spends
API budget where the teacher is actually needed: the location/reason/fix
fields of trajectories that went wrong. CORRECT trajectories are labelled for
free, because Figure 7(b) fixes their form to location="none", fix="keep".
Budget-aware by construction: Groq's free tier allows 1,000 requests/day, so
the script tracks spend on disk, stops cleanly when the quota is gone, and
resumes where it left off on the next run.
"""
from __future__ import annotations
import argparse
import json
import sys
import time
from pathlib import Path
sys.path.insert(0, str(Path(__file__).resolve().parents[1]))
if hasattr(sys.stdout, "reconfigure"):
sys.stdout.reconfigure(encoding="utf-8", errors="replace")
import src.config # noqa: F401,E402 - loads .env
from src.critic.supervise import ( # noqa: E402
Budget, BudgetExhausted, DEFAULT_TEACHER, SupervisedCritique, Teacher,
diagnose_trajectory, read_supervision, supervise_trajectory,
supervision_summary, write_supervision,
)
from src.critic.trajectory import read_trajectories # noqa: E402
def main() -> None:
ap = argparse.ArgumentParser(description=__doc__)
ap.add_argument("--trajectories", default="data/critic/trajectories_train.jsonl")
ap.add_argument("--out", default="data/critic/supervision_train.jsonl")
ap.add_argument("--budget-file", default="data/critic/.groq_budget.json")
ap.add_argument("--limit", type=int, default=None, help="max trajectories to label")
ap.add_argument("-k", "--samples", type=int, default=3,
help="teacher critiques per trajectory (Algorithm 2 K)")
ap.add_argument("--model", default=DEFAULT_TEACHER)
ap.add_argument("--max-requests", type=int, default=None,
help="cap this run's requests (default: whatever today has left)")
ap.add_argument("--mode", default="diagnose", choices=["diagnose", "consensus"],
help="diagnose: verdict is given, teacher only localises "
"(every request yields a label). consensus: Algorithm 2 "
"as published, where the teacher also votes on the verdict.")
args = ap.parse_args()
trajectories = read_trajectories(args.trajectories)
out = Path(args.out)
done = {s.qid: s for s in read_supervision(out)} if out.exists() else {}
todo = [t for t in trajectories if t.qid not in done]
if args.limit:
todo = todo[: max(0, args.limit - len(done))]
budget = Budget.load(args.budget_file)
if args.max_requests is not None:
budget.limit = min(budget.limit, budget.requests + args.max_requests)
# Only INCORRECT/UNSURE trajectories cost requests.
needs_teacher = [t for t in todo if t.verdict.value != "CORRECT"]
print(f"[sup] {len(trajectories)} trajectories, {len(done)} already labelled")
print(f"[sup] {len(todo)} to do; {len(needs_teacher)} need the teacher "
f"({len(todo) - len(needs_teacher)} derived free)")
print(f"[sup] budget today: {budget.requests}/{budget.limit} used, "
f"{budget.remaining} left -> about {budget.remaining // max(args.samples,1)} "
f"teacher-labelled trajectories at k={args.samples}")
teacher = Teacher(args.model, budget=budget)
results: list[SupervisedCritique] = list(done.values())
started, labelled, stopped = time.time(), 0, None
for i, traj in enumerate(todo, 1):
try:
label = (diagnose_trajectory if args.mode == 'diagnose'
else supervise_trajectory)
results.append(label(traj, teacher, k=args.samples))
labelled += 1
except BudgetExhausted as exc:
stopped = str(exc)
break
except Exception as exc: # noqa: BLE001 - never lose completed work
print(f"[sup] {traj.qid}: {type(exc).__name__}: {str(exc)[:90]}")
continue
if labelled % 10 == 0 or i == len(todo):
write_supervision(out, results) # checkpoint
rate = labelled / max(time.time() - started, 1e-6)
print(f"[sup] {len(results)} labelled ({rate*60:.1f}/min) "
f"budget {budget.requests}/{budget.limit}", flush=True)
write_supervision(out, results)
summary = supervision_summary(results)
print(f"\n[sup] wrote {len(results)} reference critiques to {out}")
print(json.dumps(summary, indent=2))
if stopped:
print(f"\n[sup] STOPPED: {stopped}")
print("[sup] re-run tomorrow to continue; completed work is saved.")
if __name__ == "__main__":
main()