File size: 3,583 Bytes
453b36d dcef129 453b36d dcef129 ae9e249 dcef129 ae9e249 dcef129 ae9e249 dcef129 ae9e249 dcef129 453b36d dcef129 | 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 | """Assemble the Hugging Face dataset folder for Five Desks v2 and v3 (same image; v3 trains a subset)."""
import json
import shutil
import sys
from pathlib import Path
ROOT = Path(__file__).resolve().parents[1]
sys.path.insert(0, str(ROOT))
sys.path.insert(0, str(ROOT / "scripts"))
from diligence.environment import TASK_IDS, DiligenceEnvironment, build # noqa: E402
from selection_v3 import SELECTED # noqa: E402
TIP = "put your working in the action"
# final calibration: runs a and b, with math-reasoning from the rerun after its brief was fixed
RUNS = {"a": None, "b": lambda t: not t.startswith("math-reasoning"), "reasoning2": None}
out = ROOT / "dataset"
shutil.rmtree(out, ignore_errors=True)
(out / "calibration").mkdir(parents=True)
calib = {}
for run, keep in RUNS.items():
for line in open(ROOT / f"reports/calibration-v2/{run}/episodes.jsonl"):
ep = json.loads(line)
if keep and not keep(ep["task_id"]):
continue
hinted = TIP in ep["turns"][0]["obs"]
calib.setdefault(ep["task_id"], []).append((ep["reward"], hinted))
shutil.copy(ROOT / f"reports/calibration-v2/{run}/summary.json", out / f"calibration/v2_{run}_summary.json")
shutil.copy(ROOT / f"reports/calibration-v2/{run}/episodes.jsonl", out / f"calibration/v2_{run}_episodes.jsonl")
def mean_of(pairs):
return round(sum(r for r, _ in pairs) / len(pairs), 3) if pairs else None
with open(out / "tasks.jsonl", "w") as f:
for task_id in TASK_IDS:
task = build(task_id, 0)
prompt = DiligenceEnvironment().reset(task_id=task_id, seed=0).output
f.write(json.dumps({
"task_id": task_id,
"family": task.family,
"domain": task.domain,
"level": int(task_id[-1]),
"submitted_for_training": task_id in SELECTED,
"qwen3_8_27b_mean_reward": mean_of(calib.get(task_id, [])),
"qwen3_8_27b_mean_with_hint": mean_of([p for p in calib.get(task_id, []) if p[1]]),
"qwen3_8_27b_mean_without_hint": mean_of([p for p in calib.get(task_id, []) if not p[1]]),
"qwen3_8_27b_rewards": [r for r, _ in calib.get(task_id, [])],
"example_seed": 0,
"example_prompt": prompt,
"example_files": json.dumps(task.files),
"example_solution": json.dumps(task.oracle()),
"example_caveat_missed": json.dumps(task.naive()),
"max_tool_calls": task.max_steps,
}) + "\n")
shutil.copy(ROOT / "reports/proxy/FINDINGS.md", out / "calibration/stand_in_test_findings.md")
transfer = out / "calibration/transfer"
for item in sorted((ROOT / "reports/transfer").iterdir()):
if item.is_dir() and (item / "summary.json").exists():
(transfer / item.name).mkdir(parents=True)
for name in ["summary.json", "episodes.jsonl"]:
if (item / name).exists():
shutil.copy(item / name, transfer / item.name / name)
elif item.suffix == ".txt":
transfer.mkdir(parents=True, exist_ok=True)
shutil.copy(item, transfer / item.name)
src = out / "source"
for item in ["diligence", "server", "scripts", "tests", "proxy_eval", "five_desks", "openenv.yaml",
"requirements.in", "requirements.lock", ".dockerignore"]:
path = ROOT / item
if path.is_dir():
shutil.copytree(path, src / item, ignore=shutil.ignore_patterns("__pycache__", "*.pyc"))
else:
shutil.copy(path, src / item)
shutil.copy(ROOT / "card_v3.md", out / "README.md")
print("dataset at", out, sorted(p.name for p in out.iterdir()))
|