Buckets:
| """cwm3/eval/freeze_hashes.py — the frozen-run hash table for docs/results_cwm3.md (roadmap §九): for every run dir matching the given globs, the sha256 of | |
| config.json (cfg part), base_ckpts.json, BEST checkpoint file, stats.npz, the inputs digest of run_train_*_latest.json (inputs_sha256_* listing), the layout | |
| source hashes (meta.inputs_digest per regime / split), model_id, code sha, epochs / BEST epoch. Read-only; markdown to stdout / --out. | |
| python eval/freeze_hashes.py --globs "I1_*_final" "V0_can4s_s*_stats2" "U0_kin_can_noK_s*_kcv2_clip8_ln*" "E0_pooled_P_s*_v2" "W_I1_*_final" [--out docs/results_cwm3_hashes.md] | |
| """ | |
| from __future__ import annotations | |
| import argparse | |
| import hashlib | |
| import json | |
| from pathlib import Path | |
| import numpy as np | |
| CWM3 = Path(__file__).resolve().parents[1] | |
| RUNS = CWM3 / "results" / "runs" | |
| def sha_file(p): | |
| h = hashlib.sha256() | |
| with open(p, "rb") as f: | |
| for b in iter(lambda: f.read(1 << 22), b""): | |
| h.update(b) | |
| return h.hexdigest() | |
| def sha_obj(o): | |
| return hashlib.sha256(json.dumps(o, sort_keys=True, default=str).encode()).hexdigest() | |
| def row(rd): | |
| r = {"run": rd.name} | |
| if (rd / "config.json").exists(): | |
| c = json.load(open(rd / "config.json")) | |
| r["cfg_sha"] = sha_obj(c["cfg"]) | |
| r["variant"] = c["cfg"].get("variant", "") | |
| r["kin"] = f"{c['cfg'].get('kin_source', '')}/{c['cfg'].get('kin_version', '')}" | |
| r["stats_version"] = c["cfg"].get("stats_version", "") | |
| if (rd / "base_ckpts.json").exists(): | |
| b = json.load(open(rd / "base_ckpts.json")) | |
| r["base_ckpts_sha"] = sha_obj(b.get("base_ckpts", b)) | |
| r["base_ckpts"] = b.get("base_ckpts", {}) | |
| best = (rd / "BEST").read_text().strip() if (rd / "BEST").exists() else "" | |
| r["best"] = best | |
| r["best_sha"] = sha_file(rd / best) if best and (rd / best).exists() else ("(no_train: bases as assembled)" if best == "base" else "") | |
| r["stats_sha"] = sha_file(rd / "stats.npz") if (rd / "stats.npz").exists() else "" | |
| lat = sorted(rd.glob("run_train_*_latest.json")) | |
| if lat: | |
| j = json.load(open(lat[-1])) | |
| r["inputs_digest"] = j.get("inputs", {}).get("digest", "") | |
| r["inputs_n"] = j.get("inputs", {}).get("n_files", "") | |
| r["code_sha"] = j.get("code_sha256", "") | |
| r["epochs"] = f"{j.get('epochs_run', '')} / best {j.get('best_epoch', '')}" | |
| r["run_stamp"] = j.get("run_stamp", "") | |
| lists = sorted(rd.glob("inputs_sha256_*.txt")) | |
| r["inputs_listing"] = lists[-1].name if lists else "" | |
| lay = {} | |
| for reg in ("manual", "engaged"): | |
| for sp in ("test", "val"): | |
| p = rd / f"pred_layout_{reg}_{sp}.npz" | |
| if p.exists(): | |
| m = json.loads(str(np.load(p, allow_pickle=True)["meta"])) | |
| lay[f"{reg}/{sp}"] = str(m.get("inputs_digest", ""))[:16] | |
| r["layout_src"] = lay | |
| r["done"] = ", ".join(m for m in ("TRAIN_DONE", "PRED_DONE") if (rd / m).exists()) or ("layout-only (cwm3_W warning-head layouts; no trainer markers)" if rd.name.startswith("W_") else "INCOMPLETE") | |
| if rd.name.startswith("W_"): | |
| for reg in ("manual", "engaged"): | |
| pth = rd / f"pred_layout_{reg}_test.npz" | |
| if pth.exists(): | |
| m = json.loads(str(np.load(pth, allow_pickle=True)["meta"])) | |
| r["variant"] = m.get("model_id", "")[:60] | |
| r["kin"] = "+".join(str(x.get("dir", "")).split("_final")[-1] or "base" for x in m.get("merged_from", [])) if m.get("merged_from") else "" | |
| break | |
| return r | |
| def main(): | |
| ap = argparse.ArgumentParser() | |
| ap.add_argument("--globs", nargs="+", required=True) | |
| ap.add_argument("--out", default="") | |
| a = ap.parse_args() | |
| dirs = sorted({d for g in a.globs for d in RUNS.glob(g) if d.is_dir()}) | |
| rows = [row(d) for d in dirs] | |
| lines = ["| run | variant | kin | stats v | status | BEST | BEST sha256 | cfg sha256 | base_ckpts sha256 | stats.npz sha256 | inputs digest (n files) | code sha256 | epochs | layout source sha (manual/test, engaged/test, manual/val, engaged/val) |", | |
| "|---|---|---|---|---|---|---|---|---|---|---|---|---|---|"] | |
| for r in rows: | |
| ls = r["layout_src"] | |
| lines.append(f"| {r['run']} | {r.get('variant', '')} | {r.get('kin', '')} | {r.get('stats_version', '')} | {r['done']} | {r['best']} | {str(r['best_sha'])[:16]} | {str(r.get('cfg_sha', ''))[:16]} | {str(r.get('base_ckpts_sha', ''))[:16]} | " | |
| f"{str(r['stats_sha'])[:16]} | {str(r.get('inputs_digest', ''))[:16]} ({r.get('inputs_n', '')}; {r['inputs_listing']}) | {str(r.get('code_sha', ''))[:16]} | {r.get('epochs', '')} | " | |
| + ", ".join(ls.get(k, '—') for k in ('manual/test', 'engaged/test', 'manual/val', 'engaged/val')) + " |") | |
| lines += ["", "Base checkpoints of the joint runs (base_ckpts.json):", ""] | |
| for r in rows: | |
| if r.get("base_ckpts"): | |
| lines.append(f"- `{r['run']}`: " + ", ".join(f"{b} = `{v}`" for b, v in r["base_ckpts"].items())) | |
| md = "\n".join(lines) + "\n" | |
| if a.out: | |
| Path(a.out).write_text(md) | |
| print(md) | |
| if a.out: | |
| print(f"written {a.out}") | |
| if __name__ == "__main__": | |
| main() | |
Xet Storage Details
- Size:
- 5.23 kB
- Xet hash:
- 1c12573104bbf9806cbab2f056e72eb1916c8e85b83083d8a8095809ea3efdf5
·
Xet efficiently stores files, intelligently splitting them into unique chunks and accelerating uploads and downloads. More info.