File size: 6,085 Bytes
a2b0c04 | 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 | """Beam writes the ladder setup for a community environment from its README, schema and first observation.
dotenvx run -- python scripts/setup_env.py --from-census runs/hub/sweep/census.jsonl --limit 20
For each Space it writes runs/hub/setups/<owner>__<name>.json: how many steps an episode may take, how an
episode is scored, what counts as solved and the maximum reward, each backed by a quote from the README, or
left empty when the README doesn't say. The ladder reads these files (shortlist.suite.community), so adding an
environment no longer needs a hand-written entry.
"""
from __future__ import annotations
import argparse
import json
import sys
from pathlib import Path
import requests
sys.path.insert(0, str(Path(__file__).resolve().parent.parent))
from openenv.core.generic_client import GenericEnvClient # noqa: E402
from shortlist import beam # noqa: E402
from shortlist.serve import ROOT # noqa: E402
SETUPS = ROOT / "runs" / "hub" / "setups"
SYSTEM = """You configure a test harness for a reinforcement-learning environment, using only what its README,
action schema and first observation say. Answer with one JSON object and nothing else:
{
"task": "one plain sentence: what the agent must do",
"max_steps": integer, how many actions a reasonable episode needs (at most 15),
"score": "return" if the reward is paid at the end or summed over the episode, "mean_step" if every step is scored on its own,
"max_score": number or null, the best episode score the README states or clearly implies,
"solved_at": number or null, the episode score that means the task was fully solved, if the README defines one,
"reset_kwargs": [ up to 3 objects of keyword arguments for reset() that select different tasks, e.g. {"seed": 1} or {"task_id": "easy"}; use [] if the README names none ],
"evidence": "short quotes from the README supporting max_score, solved_at and reset_kwargs, or 'none'"
}
Never guess a number the README doesn't support: use null instead."""
def safe_name(space: str) -> str:
return space.replace("/", "__")
def gather(space: str, api_url: str) -> dict:
readme = requests.get(f"https://huggingface.co/spaces/{space}/raw/main/README.md", timeout=30)
text = readme.text if readme.ok else ""
if text.startswith("---"):
text = text.split("---", 2)[-1]
schema = requests.get(api_url + "/schema", timeout=30).json().get("action", {})
with GenericEnvClient(base_url=api_url).sync() as env:
obs = env.reset().observation or {}
return {"readme": text.strip()[:12000], "schema": schema, "observation": json.dumps(obs, default=str)[:4000]}
def write_setup(space: str, api_url: str, effort: str) -> dict:
facts = gather(space, api_url)
user = (f"Environment: {space}\n\nREADME:\n{facts['readme'] or '(no README)'}\n\nAction schema:\n{json.dumps(facts['schema'])[:4000]}"
f"\n\nFirst observation after reset():\n{facts['observation']}")
response = beam.client().chat.completions.create(
model=beam.MODEL, messages=[{"role": "system", "content": SYSTEM}, {"role": "user", "content": user}],
reasoning_effort=effort, max_completion_tokens=beam.MAX_TOKENS[effort],
)
beam.USAGE.add(response.usage)
text = response.choices[0].message.content or ""
setup = None
for start in range(len(text)):
if text[start] == "{":
try:
setup = json.loads(text[start:text.rfind("}") + 1])
break
except json.JSONDecodeError:
continue
if not isinstance(setup, dict):
return {"space": space, "error": "Beam's reply had no JSON object", "reply": text[:500]}
# Keep only well-formed values; anything else falls back to the harness default.
clean = {"space": space, "api_url": api_url, "written_by": f"{beam.MODEL} ({effort} effort)",
"task": str(setup.get("task", ""))[:300], "evidence": str(setup.get("evidence", ""))[:600]}
steps = setup.get("max_steps")
clean["max_steps"] = max(1, min(15, int(steps))) if isinstance(steps, (int, float)) else 10
clean["score"] = setup.get("score") if setup.get("score") in ("return", "mean_step") else "return"
for key in ("max_score", "solved_at"):
v = setup.get(key)
clean[key] = float(v) if isinstance(v, (int, float)) and not isinstance(v, bool) else None
kwargs = setup.get("reset_kwargs")
clean["reset_kwargs"] = [k for k in kwargs if isinstance(k, dict)][:3] if isinstance(kwargs, list) else []
return clean
def main() -> None:
ap = argparse.ArgumentParser()
ap.add_argument("--from-census", default="runs/hub/sweep/census.jsonl")
ap.add_argument("--classes", default="passes floor checks,suspicious floor")
ap.add_argument("--limit", type=int, default=20)
ap.add_argument("--effort", default="low")
args = ap.parse_args()
SETUPS.mkdir(parents=True, exist_ok=True)
rows = [json.loads(l) for l in Path(args.from_census).read_text().splitlines()]
wanted = set(args.classes.split(","))
picks = sorted([r for r in rows if r["class"] in wanted and r.get("api_url")], key=lambda r: -r.get("likes", 0))[: args.limit]
for r in picks:
path = SETUPS / f"{safe_name(r['id'])}.json"
if path.exists():
continue
try:
setup = write_setup(r["id"], r["api_url"], args.effort)
except beam.DailyLimitReached:
print("Daily Beam allowance used up; run again after 00:00 UTC.")
break
except Exception as error: # noqa: BLE001
setup = {"space": r["id"], "error": f"{type(error).__name__}: {error}"[:300]}
setup["census_class"] = r["class"]
path.write_text(json.dumps(setup, indent=2))
print(f"{r['id']:55s} steps {setup.get('max_steps')} max {setup.get('max_score')} solved {setup.get('solved_at')} "
f"resets {len(setup.get('reset_kwargs', []))} {setup.get('error', '')[:80]}", flush=True)
print(json.dumps({"beam_usage": beam.USAGE.summary()}))
if __name__ == "__main__":
main()
|