Datasets:
Skills Desk v2: 36 screened training tasks (pre-registered rule), ev_contracts on CC0 USAspending data; reserved set unchanged
8685f4e verified | """Apply manifest/v2-PREREGISTRATION.md to a screen: keep candidates with 0.15 <= r <= 0.85, cap at 40 | |
| (at most 8 ev_contracts; a skill with fewer than 4 kept keeps all), order round-robin by skill. | |
| Usage: PYTHONPATH=. .venv/bin/python tools/select_v2.py runs/v2-screen-6x/records.jsonl [--trackio pooled.json] | |
| Prints a JSON report; writes nothing. --trackio: {task_id: [solved, episodes]} of v1 first-visit groups. | |
| """ | |
| import argparse | |
| import json | |
| from collections import defaultdict | |
| SKILLS = ["evidence", "compute", "transform", "repair", "constraints", "contracts"] | |
| PREFIX = {"ev": "evidence", "cmp": "compute", "tx": "transform", "fix": "repair", "plan": "constraints"} | |
| def skill(task_id: str) -> str: | |
| return "contracts" if task_id.startswith("ev_contracts") else PREFIX[task_id.split("_")[0]] | |
| def main() -> None: | |
| p = argparse.ArgumentParser() | |
| p.add_argument("records") | |
| p.add_argument("--candidates", default="manifest/v2-candidates.txt") | |
| p.add_argument("--trackio", default="") | |
| a = p.parse_args() | |
| cands = [ln.strip() for ln in open(a.candidates) if ln.strip() and not ln.startswith("#")] | |
| order = {t: i for i, t in enumerate(cands)} | |
| tally = defaultdict(lambda: [0, 0]) | |
| errors = defaultdict(int) | |
| for ln in open(a.records): | |
| r = json.loads(ln) | |
| if r["task_id"] not in order: | |
| continue | |
| if r.get("error"): | |
| errors[r["task_id"]] += 1 | |
| continue | |
| tally[r["task_id"]][0] += r["reward"] == 1.0 | |
| tally[r["task_id"]][1] += 1 | |
| if a.trackio: | |
| for t, (s, n) in json.load(open(a.trackio)).items(): | |
| tally[t][0] += s | |
| tally[t][1] += n | |
| rows = [] | |
| for t in cands: | |
| s, n = tally[t] | |
| r = s / n if n else None | |
| rows.append({"task_id": t, "skill": skill(t), "solved": s, "episodes": n, "rate": r, | |
| "errors": errors[t], "kept": r is not None and 0.15 <= r <= 0.85}) | |
| kept = sorted((x for x in rows if x["kept"]), key=lambda x: (abs(x["rate"] - 0.5), order[x["task_id"]])) | |
| by_skill = defaultdict(list) | |
| for x in kept: | |
| by_skill[x["skill"]].append(x) | |
| chosen = [] | |
| if len(kept) > 40: | |
| protected = [x for s in SKILLS for x in by_skill[s] if len(by_skill[s]) < 4] | |
| chosen = list(protected) | |
| n_contracts = sum(x["skill"] == "contracts" for x in chosen) | |
| for x in kept: | |
| if len(chosen) >= 40: | |
| break | |
| if x in chosen: | |
| continue | |
| if x["skill"] == "contracts": | |
| if n_contracts >= 8: | |
| continue | |
| n_contracts += 1 | |
| chosen.append(x) | |
| else: | |
| chosen = kept | |
| ids = {x["task_id"] for x in chosen} | |
| queues = {s: [x for x in kept if x["task_id"] in ids and x["skill"] == s] for s in SKILLS} | |
| final = [] | |
| while any(queues.values()): | |
| for s in SKILLS: | |
| if queues[s]: | |
| final.append(queues[s].pop(0)["task_id"]) | |
| print(json.dumps({"candidates": len(cands), "kept_in_band": len(kept), "selected": len(final), | |
| "freeze_allowed": len(final) >= 24, | |
| "per_skill": {s: sum(skill(t) == s for t in final) for s in SKILLS}, | |
| "order": final, "rows": rows}, indent=1)) | |
| if __name__ == "__main__": | |
| main() | |