"""Validate the published frozen release without models or extra dependencies. The recorded independent build audit used raw training/source files. This release verifier checks published bytes, schema, counts and projections; it does not rerun training-source or semantic-neighbour retrieval. """ import hashlib import json import re import unicodedata from collections import Counter from pathlib import Path ROOT = Path(__file__).resolve().parent EXPECTED_COUNTS = { "NQ": 213, "TriviaQA": 213, "PopQA": 213, "HotpotQA": 212, "2WikiMultiHopQA": 212, "MuSiQue": 212, "Bamboogle": 125, } EXPECTED_SHA256 = "ecd634aac24013213ceca3523dc024d82544c5479546fe205da24d8e8ffea678" def require(condition, message): if not condition: raise ValueError(message) def sha256(path): h = hashlib.sha256() with path.open("rb") as stream: for chunk in iter(lambda: stream.read(1024 * 1024), b""): h.update(chunk) return h.hexdigest() def normalize(question): text = " ".join(unicodedata.normalize("NFKC", question).casefold().split()) return re.sub(r"[?？]+$", "", text).strip() def jsonl(filename): rows = [] with (ROOT / filename).open(encoding="utf-8") as stream: for number, line in enumerate(stream, 1): require(bool(line.strip()), f"Blank row: {filename}:{number}") row = json.loads(line) require(isinstance(row, dict), f"Non-object row: {filename}:{number}") rows.append(row) return rows def main(): checksums = {} for line in (ROOT / "checksums.sha256").read_text().splitlines(): expected, relative = line.split(" ", 1) path = ROOT / relative require(path.resolve().is_relative_to(ROOT.resolve()), "Unsafe checksum path") require(relative not in checksums, "Duplicate checksum entry") require(sha256(path) == expected, f"SHA-256 mismatch: {relative}") checksums[relative] = expected require(checksums.get("eval_1400.jsonl") == EXPECTED_SHA256, "Unexpected frozen data SHA-256") manifest = json.loads((ROOT / "eval_1400_manifest.json").read_text()) report = json.loads((ROOT / "independent_validation.json").read_text()) audit = json.loads((ROOT / "eval_1400_exclusion_audit.json").read_text()) for value in (manifest, report, audit): require(value["status"] == "PASS", "Recorded build audit did not pass") require(manifest["total"] == report["total"] == 1400, "Wrong manifest total") require(manifest["seed"] == 42, "Wrong sampling seed") require(manifest["dataset_counts"] == report["dataset_counts"] == EXPECTED_COUNTS, "Wrong frozen composition") require(manifest["data_sha256"] == report["data_sha256"] == EXPECTED_SHA256, "Recorded data SHA-256 mismatch") for field, filename in ( ("exclusion_audit_sha256", "eval_1400_exclusion_audit.json"), ("diagnostics_sha256", "eval_1400_diagnostics.jsonl"), ("rollout_questions_sha256", "eval_1400_rollout_questions.jsonl"), ): require(manifest[field] == checksums[filename], f"Manifest mismatch: {filename}") for stage in ("SFT", "DPO", "RL"): require(manifest["exact_overlap_check"][stage] == report["training_overlaps"][stage] == audit["exact_overlap_final"][stage] == 0, f"Recorded training overlap: {stage}") require(report["all_fuzzy_flags_resolved"] is True, "Unresolved near-duplicate flags") require(report["known_high_confidence_near_duplicates"] == 0, "Recorded high-confidence near-duplicates") require(report["independent_training_exclusion_set_match"] is True, "Independent training exclusions did not match") require(report["deterministic_sampling_replay"] is True, "Sampling replay failed") require(report["rollout_gold_or_evidence_leakage"] == 0, "Recorded rollout leakage") rows = jsonl("eval_1400.jsonl") require(len(rows) == 1400, "Wrong data row count") require(Counter(row["source_dataset"] for row in rows) == EXPECTED_COUNTS, "Wrong data source counts") require([row["id"] for row in rows] == [f"eval1400_{number:04d}" for number in range(1, 1401)], "Changed IDs/order") raw, normalized = set(), set() required = {"id", "question", "answers", "source_dataset", "source_split", "source_id", "source_row_number", "source_file"} for row in rows: require(set(row) == required, f"Changed schema: {row.get('id')}") question = row["question"] require(isinstance(question, str) and bool(question.strip()), "Empty question") normal = normalize(question) require(bool(normal) and question not in raw and normal not in normalized, f"Duplicate or empty normalized question: {row['id']}") raw.add(question) normalized.add(normal) require(isinstance(row["answers"], list) and bool(row["answers"]) and all(isinstance(alias, str) and bool(alias.strip()) for alias in row["answers"]), f"Empty/invalid gold aliases: {row['id']}") source = manifest["sources"][row["source_dataset"]] require(row["source_split"] == source["source_split"] and row["source_split"] in {"dev", "test", "eval"}, "Unexpected source split") require(row["source_file"] == source["path"], "Source-file provenance mismatch") require(type(row["source_row_number"]) is int and 1 <= row["source_row_number"] <= source["expected_rows"], "Invalid source row") if row["source_dataset"] in {"NQ", "Bamboogle"}: require(row["source_id"] is None, "Unexpected source ID for an ID-less source") else: require(isinstance(row["source_id"], str) and bool(row["source_id"].strip()), "Missing original source ID") rollout = jsonl("eval_1400_rollout_questions.jsonl") require(rollout == [{"id": row["id"], "question": row["question"]} for row in rows], "Question-only rollout projection mismatch or added fields") ids = {row["id"] for row in rows} diagnostics = jsonl("eval_1400_diagnostics.jsonl") require(all(row["id"] in ids for row in diagnostics), "Unknown diagnostic ID") require(len({row["id"] for row in diagnostics}) == len(diagnostics), "Repeated diagnostic ID") print(json.dumps({"status": "PASS", "total": len(rows), "dataset_counts": dict(EXPECTED_COUNTS), "recorded_training_overlaps": report["training_overlaps"], "internal_duplicates": 0, "empty_answers": 0, "rollout_projection": "PASS", "checked_files": len(checksums), "data_sha256": EXPECTED_SHA256}, indent=2)) if __name__ == "__main__": main()