Download guard.py from holland202/tiny-engine-165k: direct link, hf CLI and curl.
- Browser
- Download file 2.69 kB
-
https://huggingface.co/holland202/tiny-engine-165k/resolve/main/guard.py
- Command line
-
hf download hf://holland202/tiny-engine-165k/guard.py
-
curl -L -o guard.py https://huggingface.co/holland202/tiny-engine-165k/resolve/main/guard.py
2.69 kB
| """Fail-closed loading of per-seed results for the registered scorers (score.py, score2.py). | |
| Added 2026-11-17 after Codex (OpenAI) showed, in draft PR #163, that the scorers counted files, not distinct seeds: | |
| five copies of one seed's result passed as five seeds. Reproduced by Claude against score2.py before this fix | |
| (five copies of seed 1 -> every criterion HELD 5/5; two of six results -> an ordinary FAILED, not an error). | |
| A result set is scored only if: the files are exactly seed1.json..seedN.json for the registered seeds; each file's | |
| "seed" field matches its file name; no two files hold the same seed or identical scores; every file's config equals | |
| the registered config; every registered score key is present and is a number in [0, 1]. Anything else exits 2 | |
| with the reason, and no verdict is printed. | |
| Written by Claude (Anthropic, Opus 5.5). SELF-TESTED only.""" | |
| import json | |
| import math | |
| import sys | |
| from pathlib import Path | |
| REGISTERED_SEEDS = (1, 2, 3, 4, 5, 6) | |
| class InvalidResults(Exception): | |
| pass | |
| def load(directory, cfg, keys, seeds=REGISTERED_SEEDS): | |
| d = Path(directory) | |
| files = sorted(d.glob("seed*.json")) | |
| names = {p.name for p in files} | |
| want = {f"seed{s}.json" for s in seeds} | |
| if names != want: | |
| raise InvalidResults(f"expected exactly {sorted(want)}; found {sorted(names)}") | |
| out, seen_scores = [], set() | |
| for s in seeds: | |
| p = d / f"seed{s}.json" | |
| try: | |
| r = json.loads(p.read_text()) | |
| except (OSError, ValueError) as e: | |
| raise InvalidResults(f"{p.name}: unreadable ({e})") from e | |
| if r.get("seed") != s: | |
| raise InvalidResults(f"{p.name}: seed field is {r.get('seed')!r}, expected {s}") | |
| if json.loads(json.dumps(cfg)) != r.get("cfg"): | |
| raise InvalidResults(f"{p.name}: config does not match the registered config") | |
| sc = r.get("scores") | |
| if not isinstance(sc, dict): | |
| raise InvalidResults(f"{p.name}: no scores") | |
| for k in keys: | |
| v = sc.get(k) | |
| if not isinstance(v, (int, float)) or isinstance(v, bool) or not math.isfinite(v) or not 0 <= v <= 1: | |
| raise InvalidResults(f"{p.name}: score {k!r} missing or not a number in [0, 1] ({v!r})") | |
| fp = json.dumps(sc, sort_keys=True) | |
| if fp in seen_scores: | |
| raise InvalidResults(f"{p.name}: scores identical to another seed's (duplicated result?)") | |
| seen_scores.add(fp) | |
| out.append(r) | |
| return out | |
| def load_or_exit(directory, cfg, keys): | |
| try: | |
| return load(directory, cfg, keys) | |
| except InvalidResults as e: | |
| print(f"INVALID RESULT SET, no verdict: {e}") | |
| sys.exit(2) | |