File size: 12,954 Bytes
cff1a8e
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
#!/usr/bin/env python3
"""Independent evidence for components/google-math, checked against the delivered artifacts.

This script deliberately does not reuse the validator's helpers beyond plain file
reads. For every sampled packed row it decodes the row's token ids back through
the SentencePiece model, then checks two things against the source record text:

1. token identity - each row slice equals the re-encoding of the corresponding
   span of the source record, taken from `[BOS] + SentencePiece(text) + [EOS]`;
2. decoded-text identity - the text decoded from the packed ids equals the text
   decoded from those source spans, modulo the whitespace normalisation that the
   shared Cayley-Math SentencePiece model applies.

It also round-trips a Parquet row to its JSONL record, and confirms a Lean record
and a generated QA record both survive intact.
"""
from __future__ import annotations

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

SPLITS = ('train', 'dev', 'test')


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


def normalise(text):
    """SentencePiece decoding collapses runs of spaces; compare on that footing."""
    return ' '.join(text.split())


def main():
    parser = argparse.ArgumentParser(description=__doc__)
    parser.add_argument('--component', type=Path, default=Path(__file__).resolve().parents[1])
    parser.add_argument('--rows', type=int, default=8, help='Packed rows to decode per split')
    args = parser.parse_args()
    root = args.component.resolve()

    import numpy as np
    import sentencepiece as spm
    import pyarrow.parquet as pq

    packed_manifest = json.loads((root / 'packed/manifest.json').read_text(encoding='utf-8'))
    tokenizer = (root / 'packed' / packed_manifest['tokenizer']['path']).resolve()
    sp = spm.SentencePieceProcessor(model_file=str(tokenizer))
    evidence = {
        'component': 'google-math',
        'tokenizer_path': str(tokenizer.relative_to(root)),
        'tokenizer_sha256': hashlib.sha256(tokenizer.read_bytes()).hexdigest(),
        'tokenizer_vocab': sp.vocab_size(),
        'packed_rows_decoded': [],
        'decoded_diffs': [],
        'parquet_roundtrips': [],
        'record_type_survival': {},
        'split_separation': {},
    }

    for split in SPLITS:
        records = read_jsonl(root / f'data/{split}.jsonl.gz')
        rows = read_jsonl(root / f'packed/{split}-rows.jsonl')
        shards = packed_manifest['splits'][split]
        by_id = {r['id']: r for r in records}
        # Sample rows spread across every shard, so a shard-only bug cannot hide, and
        # prefer rows that begin exactly at a record boundary for the literal text diff.
        boundary_rows = [i for i, r in enumerate(rows) if r['segments'][0]['record_start'] == 0]
        picks = set()
        for index in range(len(shards)):
            picks.add(min(len(rows) - 1, index * max(0, len(rows) // max(1, len(shards)))))
        step = max(1, len(rows) // args.rows)
        picks |= {min(len(rows) - 1, i * step) for i in range(args.rows)}
        boundary_picks = boundary_rows[:max(1, args.rows // 2)]
        for row_index in sorted(picks | set(boundary_picks)):
            attribute = rows[row_index]
            shard = shards[attribute['shard']]
            field = shard['fields']['input_ids']
            array = np.load(root / 'packed' / field['path'], allow_pickle=False)
            row_tokens = [int(t) for t in array[attribute['shard_row']][:attribute['length']]]
            actual_ids, expected_ids, expected_texts, mismatches = [], [], [], []
            for segment in attribute['segments']:
                record = by_id[segment['record_id']]
                reencoded = [1] + sp.encode(record['text'], out_type=int) + [2]
                span = reencoded[segment['record_start']:segment['record_end']]
                actual = row_tokens[segment['row_start']:segment['row_end']]
                if len(span) != len(actual):
                    mismatches.append({'record_id': segment['record_id'], 'reason': 'span length',
                                       'expected': len(span), 'actual': len(actual)})
                elif span != actual:
                    mismatches.append({'record_id': segment['record_id'], 'reason': 'token ids differ'})
                else:
                    actual_ids.extend(actual)
                    expected_ids.extend(span)
                    expected_texts.append(sp.decode(reencoded))
            decoded = sp.decode(actual_ids)
            record_types = sorted({by_id[s['record_id']]['record_type'] for s in attribute['segments']})
            source_ids = sorted({by_id[s['record_id']]['source_id'] for s in attribute['segments']})
            entry = {
                'split': split, 'row': row_index, 'shard': attribute['shard'],
                'shard_row': attribute['shard_row'], 'row_tokens': attribute['length'],
                'records_touched': len(attribute['segments']),
                'record_types': record_types, 'source_ids': source_ids,
                'tokens_are_exact_reencoding_of_source_spans': not mismatches,
                'token_mismatches': mismatches,
                'decoded_text_equals_source_text_modulo_whitespace':
                    normalise(decoded) == normalise(sp.decode(expected_ids)),
                'decoded_characters': len(decoded),
                'row_starts_at_record_boundary': attribute['segments'][0]['record_start'] == 0,
            }
            if entry['row_starts_at_record_boundary']:
                first = by_id[attribute['segments'][0]['record_id']]
                first_segment = attribute['segments'][0]
                source_tokens = [1] + sp.encode(first['text'], out_type=int) + [2]
                row_tokens_for_record = row_tokens[first_segment['row_start']:first_segment['row_end']]
                source_tokens_for_record = source_tokens[first_segment['record_start']:first_segment['record_end']]
                row_part = sp.decode(row_tokens_for_record)
                source_part = sp.decode(source_tokens_for_record)
                # literal line diff of the decoded prefix against the source text prefix
                diff = list(difflib.unified_diff(
                    source_part.splitlines(), row_part.splitlines(),
                    fromfile=f'source:{first["id"]}', tofile=f'packed:{split}:{row_index}', lineterm='', n=0))
                evidence['decoded_diffs'].append({
                    'split': split, 'row': row_index, 'record_id': first['id'],
                    'source_id': first['source_id'], 'record_type': first['record_type'],
                    'source_text_characters': len(first['text']),
                    'diff_line_count': len(diff),
                    'diff': diff[:40],
                    'identical_modulo_whitespace': normalise(row_part) == normalise(source_part),
                    'diff_is_whitespace_only': all(
                        line.startswith(('---', '+++', '@@')) or not line.strip() for line in diff)
                    or (normalise(row_part) == normalise(source_part)),
                    'record_first_120_chars': first['text'][:120],
                    'packed_first_120_chars': row_part[:120],
                })
            evidence['packed_rows_decoded'].append(entry)

    # Parquet round-trip: merge metadata_json with the core columns and compare to the JSONL record.
    table = pq.read_table(root / 'data/train.parquet')
    jsonl = read_jsonl(root / 'data/train.jsonl.gz')
    for index in (0, len(jsonl) // 2, len(jsonl) - 1):
        raw = table.slice(index, 1).to_pylist()[0]
        metadata = json.loads(raw.pop('metadata_json'))
        rebuilt = {**raw, **metadata}
        evidence['parquet_roundtrips'].append({
            'split': 'train', 'index': index, 'id': rebuilt['id'],
            'record_type': rebuilt['record_type'],
            'exact_match_to_jsonl_record': rebuilt == jsonl[index],
        })

    # Record-type survival: one full record of every kind, checked byte for byte.
    for split in SPLITS:
        records = read_jsonl(root / f'data/{split}.jsonl.gz')
        for record_type in ('formal_lean_proof', 'formal_lean_statement', 'generated_qa_pair',
                            'formal_geometry_problem', 'natural_language_proof', 'research_notebook'):
            match = next((r for r in records if r['record_type'] == record_type), None)
            if match is None:
                continue
            text = match['text']
            entry = {
                'split': split, 'record_type': record_type, 'id': match['id'],
                'source_id': match['source_id'], 'license': match['license'],
                'characters': len(text),
                'text_sha256_matches_record': hashlib.sha256(text.encode('utf-8')).hexdigest() == match['text_sha256'],
                'has_newlines': '\n' in text,
                'first_120_chars': text[:120],
            }
            if record_type == 'generated_qa_pair':
                entry['text_is_question_newline_answer_newline'] = text == match['question'] + '\n' + match['answer'] + '\n'
                entry['answer'] = match['answer']
            if record_type.startswith('formal_lean_'):
                entry['retains_license_header'] = ('Apache License' in text[:900] or 'Licensed under' in text[:900])
            evidence['record_type_survival'].setdefault(record_type, []).append(entry)

    # Split separation, recomputed from the delivered data rather than trusting the audit.
    group_splits = json.loads((root / 'audit/group-splits.json').read_text(encoding='utf-8'))['splits']
    seen, spanning = {}, []
    for split in SPLITS:
        for record in read_jsonl(root / f'data/{split}.jsonl.gz'):
            if record['group'] in seen and seen[record['group']] != split:
                spanning.append(record['group'])
            seen[record['group']] = split
    evidence['split_separation'] = {
        'groups_in_audit': len(group_splits),
        'groups_in_data': len(seen),
        'groups_spanning_two_splits': sorted(set(spanning)),
        'every_data_group_is_in_audit': set(seen) == set(group_splits),
        'audit_assignment_matches_data': all(group_splits[g] == s for g, s in seen.items()),
        'per_split_groups': {s: sum(1 for v in seen.values() if v == s) for s in SPLITS},
    }

    # Every declared source must actually appear in the delivered data.
    sources = json.loads((root / 'sources.json').read_text(encoding='utf-8'))['included']
    used = set()
    for split in SPLITS:
        used |= {r['source_id'] for r in read_jsonl(root / f'data/{split}.jsonl.gz')}
    evidence['sources_present'] = {'declared': sorted(sources), 'present': sorted(used),
                                  'match': sorted(used) == sorted(sources)}

    failures = [x for x in evidence['packed_rows_decoded']
                if not (x['tokens_are_exact_reencoding_of_source_spans']
                        and x['decoded_text_equals_source_text_modulo_whitespace'])]
    evidence['failures'] = failures
    evidence['rows_decoded'] = len(evidence['packed_rows_decoded'])
    evidence['splits_covered'] = sorted({x['split'] for x in evidence['packed_rows_decoded']})
    evidence['status'] = 'passed' if (
        not failures
        and len(evidence['packed_rows_decoded']) >= 3
        and all(x['exact_match_to_jsonl_record'] for x in evidence['parquet_roundtrips'])
        and not evidence['split_separation']['groups_spanning_two_splits']
        and evidence['split_separation']['every_data_group_is_in_audit']
        and evidence['split_separation']['audit_assignment_matches_data']
        and all(evidence['split_separation']['per_split_groups'][s] > 0 for s in SPLITS)
        and evidence['sources_present']['match']
        and evidence['record_type_survival'].get('formal_lean_proof')
        and evidence['record_type_survival'].get('generated_qa_pair')
    ) else 'failed'

    destination = root / 'audit/validation-evidence.json'
    destination.write_text(json.dumps(evidence, ensure_ascii=False, indent=2) + '\n', encoding='utf-8')
    print(json.dumps({k: v for k, v in evidence.items() if k != 'packed_rows_decoded'}, ensure_ascii=False, indent=2))
    print(json.dumps({'packed_rows_decoded_summary': [
        {k: x[k] for k in ('split', 'row', 'tokens_are_exact_reencoding_of_source_spans',
                           'decoded_text_equals_source_text_modulo_whitespace', 'row_tokens',
                           'record_types', 'source_ids')} for x in evidence['packed_rows_decoded']]}, indent=2))
    print(f'\nstatus: {evidence["status"]}  rows decoded: {evidence["rows_decoded"]}  '
          f'splits: {evidence["splits_covered"]}')


if __name__ == '__main__':
    main()