File size: 6,457 Bytes
ab59633 7f7e3bf ab59633 7f7e3bf ab59633 7f7e3bf ab59633 7f7e3bf ab59633 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 | """Why does the untrained model fail? Play Qwen3.8-27B on other entrants' environments and record how each episode ends.
Arena rules (strict): a reply that is not exactly one JSON object ends the episode. To see what that rule costs,
this runner keeps playing when a reply has text around a valid JSON action (lenient continuation) but records every
reply's format, so each episode yields both outcomes:
strict_reward = the reward if the episode had ended at the first non-strict reply (scored as 0 here)
lenient_reward = the reward with the format problem forgiven
Transcripts are written locally only (they contain other entrants' task text) and are never published.
Usage: HF_TOKEN_PATH=... .venv/bin/python scripts/failure_study.py --env NAME=URL --tasks a,b --samples 2 --out DIR
"""
import argparse
import json
import random
import sys
import time
import urllib.request
from concurrent.futures import ThreadPoolExecutor
from pathlib import Path
sys.path.insert(0, str(Path(__file__).resolve().parent))
sys.path.insert(0, str(Path(__file__).resolve().parents[1]))
from calibrate import Model, first_json_object, system_prompt # noqa: E402
from proxy_eval.run import classify, unwrap # noqa: E402
from wsclient import Session # noqa: E402
MAX_TURNS = 40
def play(url, model, schema, task_id, budgets, extra=""):
messages = [{"role": "system", "content": system_prompt(schema) + extra}]
ep = {"task_id": task_id, "turns": 0, "end": None, "reward": 0.0, "formats": [], "first_bad_turn": None,
"reply_tokens": 0, "obs_chars": 0}
completion, first = 0, True
with Session(url, timeout=180) as s:
data = s.reset(task_id=task_id)
while ep["turns"] < MAX_TURNS:
obs = json.dumps({k: v for k, v in data["observation"].items() if k != "metadata" or v})
messages.append({"role": "user", "content": obs})
ep["obs_chars"] += len(obs)
if not first:
completion += len(obs) // 3
first = False
if completion >= budgets[0] or sum(len(m["content"]) for m in messages) // 3 >= budgets[1]:
ep["end"] = "budget"
break
reply, used, finish_reason = model.chat(messages)
completion += used
ep["reply_tokens"] += used
messages.append({"role": "assistant", "content": reply})
ep["turns"] += 1
if finish_reason == "length":
ep["formats"].append("truncated")
ep["end"] = "truncated"
break
obj, kind = classify(reply)
fmt = kind or "clean"
if kind and (rescued := first_json_object(reply)) is not None:
obj = rescued # forgive text around a valid JSON action, but remember it
elif kind:
ep["formats"].append(fmt)
ep["end"] = fmt
break
ep["formats"].append(fmt)
if fmt != "clean" and ep["first_bad_turn"] is None:
ep["first_bad_turn"] = ep["turns"]
action = unwrap(obj)
if action.get("finish") is True:
ep["end"] = "finish"
break
try:
data = s.step(action)
except RuntimeError as exc: # the server rejected the action (schema)
ep["end"] = "schema_invalid"
ep["error"] = str(exc)[:200]
break
if data["done"]:
ep["end"], ep["reward"] = "done", float(data.get("reward") or 0.0)
break
else:
ep["end"] = "max_turns"
ep["lenient_reward"] = ep["reward"]
ep["strict_reward"] = ep["reward"] if ep["first_bad_turn"] is None and ep["end"] == "done" else 0.0
return ep, messages[1:]
def main():
ap = argparse.ArgumentParser()
ap.add_argument("--env", required=True, help="NAME=URL")
ap.add_argument("--tasks", required=True)
ap.add_argument("--samples", type=int, default=2)
ap.add_argument("--workers", type=int, default=8)
ap.add_argument("--model", default="Qwen/Qwen3.8-27B:deepinfra")
ap.add_argument("--max-tokens", type=int, default=2048)
ap.add_argument("--completion-budget", type=int, default=16384)
ap.add_argument("--context-budget", type=int, default=32768)
ap.add_argument("--system-extra", default="", help="file appended to the system prompt")
ap.add_argument("--out", required=True)
args = ap.parse_args()
extra = "\n\n" + Path(args.system_extra).read_text().strip() if args.system_extra else ""
name, url = args.env.split("=", 1)
schema = json.load(urllib.request.urlopen(url.rstrip("/") + "/schema"))["action"]
model = Model(args.model, args.max_tokens, 1.0)
jobs = [t for t in args.tasks.split(",") for _ in range(args.samples)]
random.Random(0).shuffle(jobs)
out = Path(args.out)
out.mkdir(parents=True, exist_ok=True)
started, results = time.monotonic(), []
with ThreadPoolExecutor(args.workers) as pool, open(out / "episodes.jsonl", "w") as log, \
open(out / "transcripts.jsonl", "w") as tr:
futures = [pool.submit(play, url, model, schema, t, (args.completion_budget, args.context_budget), extra)
for t in jobs]
for fut in futures:
try:
ep, messages = fut.result()
except Exception as exc: # noqa: BLE001 infrastructure errors are excluded, not scored
print("episode error:", repr(exc)[:300], file=sys.stderr)
continue
ep["env"] = name
results.append(ep)
log.write(json.dumps(ep) + "\n")
log.flush()
tr.write(json.dumps({**ep, "messages": messages}) + "\n")
n = max(1, len(results))
ends = {}
for e in results:
ends[e["end"]] = ends.get(e["end"], 0) + 1
report = {"env": name, "episodes": len(results), "seconds": round(time.monotonic() - started), "usage": model.usage,
"lenient_mean": round(sum(e["lenient_reward"] for e in results) / n, 3),
"strict_mean": round(sum(e["strict_reward"] for e in results) / n, 3),
"episodes_with_text_around_json": sum(e["first_bad_turn"] is not None for e in results), "ends": ends}
(out / "summary.json").write_text(json.dumps(report, indent=2) + "\n")
print(json.dumps(report))
if __name__ == "__main__":
main()
|