Datasets:
Publish cross-hardware reproducibility dataset with preprint PDF and source (CC-BY-4.0)
4e189ca verified | #!/usr/bin/env python3 | |
| """Rebuild the paper's inputs from the PUBLISHED dataset alone, then regenerate every number. | |
| python3 figures/from_dataset.py --dataset <root of the HF dataset> --work <scratch dir> | |
| python3 figures/make_figures.py --decl <work>/decl-nemo --decl <work>/decl-batch2 --items <work>/items | |
| The dataset stores the three runs of each item side by side (data/items.jsonl). This script splits | |
| them back into one items.jsonl per run (<work>/items/runs/<run_id>/items.jsonl), and sorts the | |
| runtime declarations into the two batch directories that make_figures.py expects (batch 1 is the | |
| directory whose name contains "nemo"). make_figures.py then rebuilds every card, cross-checks it | |
| against the declarations the signed capsules bind (a mismatch aborts), and writes build/numbers.json | |
| and tables/*.tex; compare them with the released copies. | |
| """ | |
| from __future__ import annotations | |
| import argparse | |
| import json | |
| import os | |
| import shutil | |
| RUNS = {"rtx3090": "rtx3090_", "t4_run1": "t4_run1_", "t4_run2": "t4_run2_"} | |
| FIELDS = ("raw_output", "raw_output_sha256", "parsed_label", "grade", "done_reason") | |
| def main() -> None: | |
| ap = argparse.ArgumentParser() | |
| ap.add_argument("--dataset", required=True) | |
| ap.add_argument("--work", required=True) | |
| a = ap.parse_args() | |
| ds, wk = a.dataset, a.work | |
| if os.path.exists(wk): | |
| raise SystemExit(f"{wk} exists; pass a fresh directory") | |
| withheld = [l for l in open(os.path.join(ds, "data/withheld.jsonl"), encoding="utf-8") if l.strip()] | |
| if withheld: | |
| raise SystemExit(f"{len(withheld)} outputs are withheld from this copy; the prefix analysis needs them") | |
| batch_of = {} | |
| for line in open(os.path.join(ds, "data/cards.jsonl"), encoding="utf-8"): | |
| c = json.loads(line) | |
| batch_of[os.path.basename(c["declaration_file"])] = c["batch"] | |
| for b, name in (("batch1", "decl-nemo"), ("batch2", "decl-batch2")): | |
| os.makedirs(os.path.join(wk, name)) | |
| for f, cb in batch_of.items(): | |
| if cb == b: | |
| shutil.copyfile(os.path.join(ds, "declarations", f), os.path.join(wk, name, f)) | |
| runs: dict[str, list[dict]] = {} | |
| for line in open(os.path.join(ds, "data/items.jsonl"), encoding="utf-8"): | |
| r = json.loads(line) | |
| for pre in RUNS.values(): | |
| rid = r.get(pre + "run_id") | |
| if rid is None: | |
| continue | |
| rec = {"item_id": r["item_id"], "sequence": r["sequence"], "prompt_sha256": r["prompt_sha256"], | |
| "predicate": r["predicate"], "transport_ok": r[pre + "status"] != "transport_error"} | |
| rec.update({k: r[pre + k] for k in FIELDS}) | |
| runs.setdefault(rid, []).append(rec) | |
| for rid, recs in runs.items(): | |
| d = os.path.join(wk, "items", "runs", rid) | |
| os.makedirs(d) | |
| with open(os.path.join(d, "items.jsonl"), "w", encoding="utf-8") as fh: | |
| for rec in sorted(recs, key=lambda x: x["sequence"]): | |
| fh.write(json.dumps(rec, ensure_ascii=False) + "\n") | |
| print(f"{len(batch_of)} declarations, {len(runs)} runs written under {wk}") | |
| if __name__ == "__main__": | |
| main() | |