| """Assemble the Hugging Face dataset folder for Five Desks v5: the finance specialist or the collapse test. |
| |
| Usage: .venv/bin/python scripts/build_dataset_v5.py --mode finance|control (writes dataset_v5_<mode>/) |
| """ |
|
|
| import argparse |
| import glob |
| import json |
| import shutil |
| import sys |
| from pathlib import Path |
| from statistics import mean |
|
|
| ROOT = Path(__file__).resolve().parents[1] |
| sys.path.insert(0, str(ROOT)) |
|
|
| from diligence.environment import CONTROL_PREFIX, DiligenceEnvironment, build |
|
|
|
|
| def calibration(pattern): |
| rewards, lengths = {}, {} |
| for f in glob.glob(str(ROOT / pattern)): |
| for line in open(f): |
| ep = json.loads(line) |
| rewards.setdefault(ep["task_id"], []).append(ep["reward"]) |
| lengths.setdefault(ep["task_id"], []).append(len(ep["turns"])) |
| return rewards, lengths |
|
|
|
|
| def main(): |
| ap = argparse.ArgumentParser() |
| ap.add_argument("--mode", choices=["finance", "control"], required=True) |
| args = ap.parse_args() |
| request = json.loads((ROOT / f"submission_v5_{args.mode}_candidate.json").read_text()) |
| out = ROOT / f"dataset_v5_{args.mode}" |
| shutil.rmtree(out, ignore_errors=True) |
| (out / "calibration").mkdir(parents=True) |
|
|
| a_rewards, a_lengths = calibration("reports/calibration-v4a/*/episodes.jsonl") |
| b_rewards, _ = calibration("reports/calibration-code/*/episodes.jsonl") |
| with open(out / "tasks.jsonl", "w") as f: |
| for task in request["tasks"]: |
| tid = task["task_id"] |
| row = {"task_id": tid, "image": task.get("image", request["image"]), "submitted_for_training": True} |
| base = tid.removeprefix(CONTROL_PREFIX) |
| if tid.startswith("bugfix-"): |
| r = b_rewards.get(tid, []) |
| row.update(qwen3_8_27b_mean_reward=round(mean(r), 3) if r else None, qwen3_8_27b_rewards=r) |
| else: |
| t = build(base, 0) |
| row.update(family=t.family, domain=t.domain, max_tool_calls=t.max_steps, |
| example_prompt=DiligenceEnvironment().reset(task_id=tid, seed=0).output) |
| if tid.startswith(CONTROL_PREFIX): |
| calls = a_lengths.get(base, []) |
| row.update(untrained_calls_per_episode=calls, |
| untrained_control_reward=round(mean((t.max_steps - c + 1) / t.max_steps for c in calls), 3) |
| if calls else None) |
| else: |
| r = a_rewards.get(tid, []) |
| row.update(qwen3_8_27b_mean_reward=round(mean(r), 3) if r else None, qwen3_8_27b_rewards=r) |
| f.write(json.dumps(row) + "\n") |
|
|
| for run in sorted(glob.glob(str(ROOT / "reports/calibration-v4a/*"))) + ( |
| sorted(glob.glob(str(ROOT / "reports/calibration-code/*"))) if args.mode == "finance" else []): |
| tag = Path(run).parent.name.replace("calibration-", "") + "_" + Path(run).name |
| for name in ["summary.json", "episodes.jsonl"]: |
| if (Path(run) / name).exists(): |
| shutil.copy(Path(run) / name, out / "calibration" / f"{tag}_{name}") |
| shutil.copy(ROOT / "reports/lottery-analysis-2026-10-12.md", out / "calibration/lottery-analysis.md") |
| shutil.copy(ROOT / "reports/v4-arena-gap.md", out / "calibration/v4-arena-gap.md") |
|
|
| src = out / "source" |
| skip = shutil.ignore_patterns("__pycache__", "*.pyc", "data", "out", "sdists", "*.log") |
| for item in ["diligence", "codedesk", "server", "scripts", "tests", "proxy_eval", "openenv.yaml", |
| "requirements.in", "requirements.lock", ".dockerignore", "CHANGELOG.md"]: |
| path = ROOT / item |
| if path.is_dir(): |
| shutil.copytree(path, src / item, ignore=skip) |
| elif path.exists(): |
| shutil.copy(path, src / item) |
| shutil.copy(ROOT / f"card_v5_{args.mode}.md", out / "README.md") |
| if args.mode == "control": |
| shutil.copy(ROOT / "submission_v5_control_candidate.json", out / "submission_request.json") |
| print("dataset at", out, sorted(p.name for p in out.iterdir())) |
|
|
|
|
| if __name__ == "__main__": |
| main() |
|
|
