ATOM (atom-ai): code snapshot (100fae0 + batched-train), README with measured numbers, curated checkpoints (part 3)
969c687 verified | """Append one progress row for a mixed-corpus checkpoint: held-out dialogue / held-out varied / old | |
| val2048 (normal + field-alone), train stats, 3 probe chats (unseeded) + seed 1. 1 thread. | |
| usage: tsv_mix.py TSV CKPT RUN_METRICS_JSON [--no-chats]""" | |
| import json, os, re, sys, time | |
| from pathlib import Path | |
| import torch | |
| sys.path.insert(0, str(Path(__file__).resolve().parent)) | |
| import mix_common as M | |
| def main(): | |
| torch.set_num_threads(1) | |
| tsv, ck, rm = sys.argv[1:4] | |
| t0 = time.time(); m = M.C.load_model(ck); v = M.all_vals(m) | |
| tr = json.load(open(rm)).get("train", {}) if os.path.exists(rm) else {} | |
| lam = tr.get("field_leak_learnable_rates") | |
| chats = [] if "--no-chats" in sys.argv else M.C.probe_chats(m, seeds=(1,)) | |
| step = re.search(r"step_(\d+)_", ck).group(1) | |
| cols = [("step", step)] + [(k, f"{x:.4f}") for k, x in v.items()] + [ | |
| ("train_final_loss", f"{tr.get('final_loss', float('nan')):.4f}"), | |
| ("steps_per_s", f"{tr.get('transitions_per_second', 0):.1f}"), | |
| ("cap_frac", f"{tr.get('field_limited_fraction', float('nan')):.3f}"), | |
| ("nan", str(tr.get("nan_or_inf"))), | |
| ("leak_mean", f"{sum(lam) / len(lam):.4f}" if lam else ""), | |
| ("leak_min", f"{min(lam):.4f}" if lam else ""), | |
| ("time", time.strftime("%Y-%m-%dT%H:%M:%S%z"))] | |
| cols += [(f"chat_{r['seed'] or 'u'}_{i % 3}", " ".join(r["resp"].split())[:90].replace("\t", " ")) for i, r in enumerate(chats)] | |
| new = not os.path.exists(tsv) | |
| with open(tsv, "a", encoding="utf-8") as fh: | |
| if new: fh.write("\t".join(c for c, _ in cols) + "\n") | |
| fh.write("\t".join(x for _, x in cols) + "\n") | |
| print("\t".join(x for _, x in cols), f"({time.time() - t0:.0f}s)", flush=True) | |
| if __name__ == "__main__": | |
| main() | |