File size: 22,311 Bytes
5f79a08
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
#!/usr/bin/env python3
"""Validate components/google-math: file integrity, provenance, leakage controls, packed tokens."""
from __future__ import annotations

import argparse
import gzip
import hashlib
import json
import math
from pathlib import Path

CORE = ('id', 'text', 'group', 'license', 'paper_id', 'split', 'record_type', 'source_url')
SPLITS = ('train', 'dev', 'test')
RECORD_TYPES = {'formal_lean_proof', 'natural_language_proof', 'formal_lean_statement',
                'formal_geometry_problem', 'geometry_rule_table', 'generated_qa_pair',
                'research_notebook'}
VERBATIM_QUALITY = {'verbatim_source_text'}


def ensure(condition, message):
    if not condition:
        raise ValueError(message)


def sha(path):
    h = hashlib.sha256()
    with Path(path).open('rb') as f:
        for chunk in iter(lambda: f.read(1024 * 1024), b''):
            h.update(chunk)
    return h.hexdigest()


def read_jsonl(path):
    path = Path(path)
    opener = gzip.open if path.suffix == '.gz' else open
    with opener(path, 'rt', encoding='utf-8') as f:
        return [json.loads(line) for line in f if line.strip()]


def read_json(path):
    return json.loads(Path(path).read_text(encoding='utf-8'))


def safe_path(root, relative):
    ensure(not Path(relative).is_absolute(), f'Absolute manifest path: {relative}')
    p = (root / relative).resolve()
    ensure(p.is_relative_to(root.resolve()), f'Path escapes component root: {relative}')
    return p


def check_parquet(path, expected):
    import pyarrow.parquet as pq
    table = pq.read_table(path)
    ensure(table.column_names == [*CORE, 'metadata_json'], f'Parquet columns: {path}')
    ensure(table.num_rows == len(expected), f'Parquet row count: {path}')
    for index, (row, original) in enumerate(zip(table.to_pylist(), expected)):
        metadata = json.loads(row.pop('metadata_json'))
        ensure({**row, **metadata} == original, f'Parquet mismatch {path}:{index}')


def validate_packed(root, split_records):
    import numpy as np
    import sentencepiece as spm
    packed = root / 'packed'
    manifest = read_json(packed / 'manifest.json')
    ensure(manifest['format'] == 'mmla-type2-packed-v1', 'Packed format')
    ensure(manifest['seq_len'] == 2048, 'Packed sequence length')
    ensure(manifest['source'] == 'google-math', 'Packed source')
    tokenizer = (packed / manifest['tokenizer']['path']).resolve()
    ensure(tokenizer.is_relative_to(root), 'Tokenizer path escapes component')
    ensure(sha(tokenizer) == manifest['tokenizer']['sha256'], 'Tokenizer hash')
    sp = spm.SentencePieceProcessor(model_file=str(tokenizer))
    ensure((sp.pad_id(), sp.bos_id(), sp.eos_id(), sp.unk_id(), sp.vocab_size()) == (0, 1, 2, 3, 32768),
           'Tokenizer IDs')
    result = {}
    for split in SPLITS:
        records = split_records[split]
        token_catalog = read_jsonl(packed / f'{split}-records.jsonl')
        rows = read_jsonl(packed / f'{split}-rows.jsonl')
        ensure([r['id'] for r in records] == [r['id'] for r in token_catalog], f'{split} packed source ordering')
        arrays, doc_offsets, stream_start = [], [], 0
        for r, entry in zip(records, token_catalog):
            tokens = np.asarray([1] + sp.encode(r['text'], out_type=int) + [2], dtype='<i4')
            ensure(len(tokens) == entry['tokens'], f'{split} record token count')
            ensure(hashlib.sha256(tokens.tobytes()).hexdigest() == entry['token_sha256'],
                   f'{split} record token hash')
            ensure(r['paper_id'] == entry['paper_id'], f'{split} record paper ID')
            doc_offsets.append((stream_start, stream_start + len(tokens), r['id'], r['paper_id']))
            stream_start += len(tokens)
            arrays.append(tokens)
        stream = np.concatenate(arrays)
        del arrays
        cursor, doc_index, row_number, labels_count = 0, 0, 0, 0
        for shard_index, shard in enumerate(manifest['splits'][split]):
            loaded = {}
            for name in ('input_ids', 'labels', 'role_ids'):
                field = shard['fields'][name]
                path = safe_path(packed, field['path'])
                ensure(path.stat().st_size == field['bytes'] and sha(path) == field['sha256'],
                       f'Packed file hash/size: {path}')
                array = np.load(path, mmap_mode='r', allow_pickle=False)
                ensure(array.dtype == np.dtype('int32'), 'Packed dtype')
                ensure(list(array.shape) == field['shape'] == [shard['count'], 2048], 'Packed shape')
                loaded[name] = array
            ensure(bool(np.all(loaded['role_ids'] == 0)), 'Nonzero plain-text role')
            ensure(int(loaded['input_ids'].min()) >= 0, 'Negative input token')
            for shard_row in range(shard['count']):
                ensure(row_number < len(rows), 'Missing row attribution')
                attribution = rows[row_number]
                n = min(2048, len(stream) - cursor)
                ensure(n > 0, 'Unexpected all-padding row')
                expect_input = np.zeros(2048, dtype=np.int32)
                expect_input[:n] = stream[cursor:cursor + n]
                expect_labels = np.full(2048, -100, dtype=np.int32)
                available = min(n, len(stream) - cursor - 1)
                if available > 0:
                    expect_labels[:available] = stream[cursor + 1:cursor + 1 + available]
                expect_labels[expect_labels == 1] = -100
                ensure(np.array_equal(loaded['input_ids'][shard_row], expect_input),
                       f'{split} source token mismatch at row {row_number}')
                ensure(np.array_equal(loaded['labels'][shard_row], expect_labels),
                       f'{split} label shift/mask mismatch at row {row_number}')
                ensure(attribution['row'] == row_number and attribution['shard'] == shard_index
                       and attribution['shard_row'] == shard_row, 'Row location attribution')
                ensure(attribution['stream_start'] == cursor and attribution['length'] == n,
                       'Row stream attribution')
                ensure(attribution['segments'], 'Row without a source segment')
                while doc_index < len(doc_offsets) and doc_offsets[doc_index][1] <= cursor:
                    doc_index += 1
                spans, k = [], doc_index
                while k < len(doc_offsets) and doc_offsets[k][0] < cursor + n:
                    start, end, rid, pid = doc_offsets[k]
                    lo, hi = max(cursor, start), min(cursor + n, end)
                    if lo < hi:
                        spans.append({'record_id': rid, 'paper_id': pid,
                                      'row_start': lo - cursor, 'row_end': hi - cursor,
                                      'record_start': lo - start, 'record_end': hi - start})
                    k += 1
                ensure(attribution['segments'] == spans, 'Token source attribution mismatch')
                labels_count += int(np.count_nonzero(expect_labels != -100))
                cursor += n
                row_number += 1
        ensure(cursor == len(stream), f'{split} token stream not fully consumed')
        ensure(row_number == len(rows), f'{split} extra row attribution')
        actual = {'rows': row_number, 'tokens': len(stream), 'loss_bearing_labels': labels_count}
        ensure(actual == manifest['statistics'][split], f'{split} packed statistics mismatch')
        result[split] = actual
        print(f'Validated packed {split}: {row_number} rows, {len(stream)} tokens', flush=True)
    return result


def main():
    parser = argparse.ArgumentParser(description=__doc__)
    parser.add_argument('--component', type=Path, default=Path(__file__).resolve().parents[1])
    parser.add_argument('--source-root', type=Path,
                        help='Pinned upstream checkouts to hash-check source files against')
    args = parser.parse_args()
    root = args.component.resolve()
    checks = []

    manifest = read_json(root / 'manifest.json')
    ensure(manifest['format'] == 'cayley-contribution-file-manifest-v1', 'Manifest format')
    ignored = set(manifest['excludes'])
    expected_files = {x['path'] for x in manifest['files']}
    actual_files = {p.relative_to(root).as_posix() for p in root.rglob('*')
                    if p.is_file() and p.relative_to(root).as_posix() not in ignored}
    ensure(expected_files == actual_files,
           f'File manifest coverage mismatch: missing={sorted(expected_files - actual_files)}, '
           f'extra={sorted(actual_files - expected_files)}')
    ensure(not any('__pycache__' in p or p.endswith('.pyc') for p in actual_files), 'Build artefacts in component')
    ensure(not any('.DS_Store' in p for p in actual_files), '.DS_Store in component')
    for item in manifest['files']:
        path = safe_path(root, item['path'])
        ensure(path.stat().st_size == item['bytes'], f'Byte count: {path}')
        ensure(sha(path) == item['sha256'], f'File digest: {path}')
    checks.append('all_delivered_file_hashes_and_sizes')

    sources = read_json(root / 'sources.json')
    ensure(sources['mathematical_verification'] == 'not_performed', 'sources.json verification note')
    ensure(sources['all_records_license'] == 'per-record', 'sources.json license mode')
    included = sources['included']
    ensure(len(included) >= 5, 'Fewer than five sources')
    for source_id, source in included.items():
        ensure(source['license'] in {'Apache-2.0', 'CC-BY-4.0', 'MIT'}, f'{source_id} redistribution license')
        ensure(bool(source['commit']), f'{source_id} missing pinned revision')
        ensure(source['fixed_revision_url'].endswith(source['commit']), f'{source_id} fixed revision URL')
        ensure(source['license_file'] == f'licenses/{source_id}/LICENSE', f'{source_id} license index')
        ensure((root / source['license_file']).is_file(), f'{source_id} license file missing')
    ensure(len(sources['excluded_candidates']) >= 8, 'Too few excluded candidates documented')
    for candidate in sources['excluded_candidates']:
        ensure(candidate['reason'].strip(), 'Excluded candidate without a reason')
    checks.append('pinned_sources_metadata_and_excluded_candidates')

    source_files = read_jsonl(root / 'audit/source-files.jsonl')
    ensure(source_files, 'Empty source file inventory')
    ensure(len({(x['source_id'], x['path']) for x in source_files}) == len(source_files),
           'Duplicate source inventory entry')
    ensure({x['source_id'] for x in source_files} == set(included), 'Source inventory coverage')
    ensure(all(x['kind'] in {'upstream_blob', 'generated_by_pinned_generator'} for x in source_files),
           'Source inventory kind')
    if args.source_root:
        pinned = args.source_root.resolve()
        for item in source_files:
            if item['kind'] != 'upstream_blob':
                continue
            path = safe_path(pinned / SOURCES_DIR[item['source_id']], item['path'])
            ensure(path.is_file(), f'Missing upstream file: {item["path"]}')
            ensure(path.stat().st_size == item['bytes'], f'Upstream size: {item["path"]}')
            ensure(sha(path) == item['sha256'], f'Upstream digest: {item["path"]}')
    checks.append('source_file_inventory' + ('_local_bytes' if args.source_root else '_metadata_only'))

    group_splits = read_json(root / 'audit/group-splits.json')
    assigned = group_splits['splits']
    ensure(set(assigned.values()) == set(SPLITS), f'Splits present: {set(assigned.values())}')
    ensure(len(assigned) >= 3, 'Too few groups for leakage-safe splits')
    for group in assigned:
        ensure('::' in group, f'Group not namespaced by source: {group}')
    # The split policy is fully determined by the group set, so recompute it from scratch.
    ordered = sorted(assigned, key=lambda g: (hashlib.sha256(g.encode()).hexdigest(), g))
    holdout = max(1, math.ceil(len(ordered) * .01))
    if 2 * holdout >= len(ordered):
        holdout = 1
    recomputed = {g: ('dev' if i < holdout else 'test' if i < 2 * holdout else 'train')
                  for i, g in enumerate(ordered)}
    ensure(recomputed == assigned, 'Severity: recorded group assignment differs from the stated policy')
    checks.append('group_split_policy_recomputed_from_scratch')

    split_records, ids, hashes, groups_seen, sources_seen = {}, set(), set(), {}, set()
    source_by_key = {(x['source_id'], x['path']): x for x in source_files}
    for split in SPLITS:
        rows = read_jsonl(root / f'data/{split}.jsonl.gz')
        ensure(rows, f'Empty split {split}')
        check_parquet(root / f'data/{split}.parquet', rows)
        for r in rows:
            ensure(r['split'] == split and assigned[r['group']] == split, 'Group split assignment')
            ensure(r['record_type'] in RECORD_TYPES, f'Record type: {r["id"]}')
            ensure(r['license'] == included[r['source_id']]['license'], 'Record license matches source index')
            ensure(r['source_url'].startswith(included[r['source_id']]['repository'] + '/'), 'Record source URL host')
            ensure(included[r['source_id']]['commit'] in r['source_url'], 'Record pinned revision')
            ensure(r['mathematical_verification'] == 'not_performed' and r['lean_verified'] is False,
                   'Verification claim')
            ensure(r['creators'] == included[r['source_id']]['creators'], 'Record creators')
            ensure(r['group'].startswith(r['source_id'] + '::'), 'Group not namespaced by source_id')
            ensure(r['id'] not in ids, 'Cross-split duplicate ID')
            ids.add(r['id'])
            digest = hashlib.sha256(r['text'].encode('utf-8')).hexdigest()
            ensure(r['text_sha256'] == digest and digest not in hashes, 'Exact text duplicate or digest mismatch')
            hashes.add(digest)
            ensure(r['text'].strip(), f'Empty text: {r["id"]}')
            if r['record_type'] == 'generated_qa_pair':
                ensure(r['content_origin'] == 'generator_output', 'Generator origin label')
                ensure(r['text'] == r['question'] + '\n' + r['answer'] + '\n', 'Generated text is question+answer')
                ensure('Question' not in r['question'][:12], 'Generated question rewritten')
                files_entry = source_by_key.get((r['source_id'], 'generated/mathematics_dataset_sample.jsonl'))
                ensure(files_entry is not None, 'Generated corpus file absent from inventory')
            else:
                files_entry = source_by_key[(r['source_id'], r['source_path'])]
                ensure(files_entry['sha256'] == r['source_sha256'], 'Record source digest mismatch')
            if r['record_type'] in {'formal_lean_proof', 'formal_lean_statement'}:
                ensure(r['extraction_quality'] in VERBATIM_QUALITY, 'Lean extraction quality')
                ensure(r['transform'] == 'identity_verbatim_bytes', 'Lean transform')
                if args.source_root:
                    original = safe_path(args.source_root.resolve() / SOURCES_DIR[r['source_id']],
                                         r['source_path']).read_text(encoding='utf-8')
                    ensure(r['text'] == original, f'Lean text differs from pinned bytes: {r["source_path"]}')
            elif r['record_type'] == 'natural_language_proof':
                ensure(r['extraction_quality'] == 'lossy_pdf_text_layer', 'PDF extraction quality flag')
                ensure(r['extraction_metrics']['pages'] >= 1, 'PDF page count')
                ensure(r['extraction_metrics']['characters'] == len(r['text']), 'PDF character count')
            elif r['record_type'] in {'formal_geometry_problem', 'geometry_rule_table',
                                      'research_notebook'}:
                ensure(r['extraction_quality'] in {'verbatim_source_text', 'verbatim_cell_text'},
                       'Text artifact quality flag')
            if r['group'] in groups_seen:
                ensure(groups_seen[r['group']] == split, 'Group cross-split leakage')
            groups_seen[r['group']] = split
            sources_seen.add(r['source_id'])
        split_records[split] = rows
    ensure(sources_seen == set(included), f'Sources absent from data: {sorted(set(included) - sources_seen)}')
    checks.append('lossless_jsonl_parquet_exact_text_uniqueness_and_record_provenance')
    checks.append('group_split_separation')
    checks.append('verbatim_lean_source_equality' + ('_local_bytes' if args.source_root else '_metadata_only'))

    # A Lean source and a generated QA record must both be present and both non-trivial.
    lean_ids = [r['id'] for r in sum(split_records.values(), []) if r['record_type'] == 'formal_lean_proof']
    generated_ids = [r['id'] for r in sum(split_records.values(), []) if r['record_type'] == 'generated_qa_pair']
    ensure(lean_ids, 'No Lean proof record survived')
    ensure(len(generated_ids) > 1000, 'Too few generated QA records survived')
    checks.append('lean_and_generated_records_survive')

    extraction = read_jsonl(root / 'audit/extraction.jsonl')
    ensure(extraction, 'Empty extraction audit')
    prose = {r['source_path']: r for r in sum(split_records.values(), [])
             if r['record_type'] == 'natural_language_proof'}
    prose_audit = [x for x in extraction if x['source_id'] == 'alphaproof-nexus']
    ensure({x['source_path'] for x in prose_audit if x.get('pages')} == set(prose),
           'Extraction audit coverage')
    for entry in prose_audit:
        if not entry.get('pages'):
            continue
        ensure(entry['extraction_quality'] == 'lossy_pdf_text_layer', 'Extraction audit quality flag')
        ensure(len(entry['page_characters']) == entry['pages'], 'Per-page counts')
        ensure(entry['characters'] == prose[entry['source_path']]['extraction_metrics']['characters'],
               'Extraction character count')
    checks.append('per_file_extraction_audit')

    duplicates = read_jsonl(root / 'audit/excluded-duplicates.jsonl')
    ensure(len({d['excluded_id'] for d in duplicates}) == len(duplicates), 'Duplicate excluded IDs')
    checks.append('dedup_audit_present')

    generation = read_json(root / 'audit/mathematics-dataset-generation.json')
    ensure(generation['generator_commit'] == included['mathematics-dataset']['commit'], 'Generator commit pin')
    generated_duplicates = sum(1 for d in duplicates if d['source_id'] == 'mathematics-dataset')
    ensure(generation['measured']['records'] - generated_duplicates == len(generated_ids),
           'Generator audit record count after exact deduplication')
    ensure(generation['parameters']['seed'] == generation['parameters']['seed'], 'Generator seed recorded')
    ensure(generation['parameters']['generator_commit'] is None or
           generation['parameters']['generator_commit'] in {'', None} or
           generation['parameters']['generator_commit'] == included['mathematics-dataset']['commit'],
           'Generator commit in parameters')
    checks.append('generator_audit_seed_and_quota')

    packed = validate_packed(root, split_records) if (root / 'packed/manifest.json').exists() else {}
    if packed:
        checks.append('all_packed_tokens_reencoded_source_shift_padding_masks_and_attribution')

    stats = read_json(root / 'statistics.json')
    ensure(stats['records'] == sum(len(v) for v in split_records.values()), 'Statistics record count')
    ensure(stats['records'] == stats['records_before_deduplication'] - stats['excluded_exact_duplicates'],
           'Statistics dedup accounting')
    ensure(stats['groups'] == len(assigned), 'Statistics group count')
    ensure(stats['packed'] == packed, 'Statistics packed counts')
    ensure(stats['text_characters'] == sum(len(r['text']) for v in split_records.values() for r in v),
           'Statistics character count')
    for source_id, source_stats in stats['per_source'].items():
        actual = sum(1 for v in split_records.values() for r in v if r['source_id'] == source_id)
        ensure(source_stats['records'] == actual, f'Per-source record count {source_id}')
        ensure(source_stats['text_characters'] == sum(len(r['text']) for v in split_records.values()
                                                      for r in v if r['source_id'] == source_id),
               f'Per-source character count {source_id}')
    for split in SPLITS:
        ensure(stats['splits'][split]['records'] == len(split_records[split]), f'Statistics {split} records')
    checks.append('statistics_match_artifacts')

    report = {
        'status': 'passed',
        'component': 'google-math',
        'sources': len(included),
        'records': stats['records'],
        'groups': stats['groups'],
        'record_types': stats['record_types'],
        'splits': stats['splits'],
        'source_files_hashed': len(source_files),
        'packed': packed,
        'checks': checks,
        'mathematical_verification': 'not_performed',
        'lean_compilation': 'not_performed',
        'limitations': [
            'No Lean compilation, proof checking or mathematical verification.',
            'PDF extraction is lossy for symbols, ligatures and formula layout.',
            'Generated QA pairs are a seeded sample of a school-level curriculum, not research mathematics '
            'and not the paper-complete dataset.',
            'formal-conjectures upstream notes third-party CC-BY-SA derived content; no per-file license audit.',
            'No near-duplicate or Phase 1 overlap assessment.',
            'SentencePiece decoding normalises runs of spaces, inherited from the shared Cayley-Math tokenizer.',
        ],
    }
    (root / 'validation.json').write_text(json.dumps(report, ensure_ascii=False, indent=2) + '\n', encoding='utf-8')
    print(json.dumps(report, ensure_ascii=False, indent=2))


SOURCES_DIR = {
    'alphaproof-nexus': 'alphaproof-nexus-results',
    'formal-conjectures': 'formal-conjectures',
    'formal-imo': 'formal-imo',
    'formal-putnam-like': 'formal-putnam-like',
    'cryptids': 'cryptids',
    'mathematics-dataset': 'mathematics_dataset',
    'alphageometry': 'alphageometry',
    'mathematics-conjectures': 'mathematics_conjectures',
}


if __name__ == '__main__':
    main()