Datasets:
File size: 8,295 Bytes
7a088d4 | 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 | """Create corpus/token derivatives excluding the audited unknown-era Kanripo IDs."""
import argparse
from collections import Counter
import hashlib
import json
from pathlib import Path
import shutil
import numpy as np
from sentencepiece_unigram import now, read_json, sha, write_json
def excluded(row):
return row['source']=='kanripo' and row['date_evidence'].get('dynasty')=='時代不明'
def main():
p=argparse.ArgumentParser(description=__doc__)
p.add_argument('--input',type=Path,default=Path('unified/pretrain-v1/build-002'))
p.add_argument('--tokens',type=Path,default=Path('unified/tokens-v1/build-001'))
p.add_argument('--audit',type=Path,default=Path('kanripo/unknown-era-audit/run-001'))
p.add_argument('--output',type=Path,default=Path('unified/pretrain-v1/build-003'))
p.add_argument('--token-output',type=Path,default=Path('unified/tokens-v1/build-002'))
a=p.parse_args()
cm,tm,am=(read_json(a.input/'manifest.json'),read_json(a.tokens/'manifest.json'),read_json(a.audit/'report.json'))
if not all(m['complete'] for m in (cm,tm,am)):
raise ValueError('Incomplete inputs')
src=a.input/'pretrain.jsonl'
if sha(src)!=cm['output_sha256'][src.name] or tm['input_sha256']!=cm['output_sha256'][src.name] or am['input_sha256']!=tm['input_sha256']:
raise ValueError('Input identity mismatch')
for name,expected in tm['files'].items():
if sha(a.tokens/name)!=expected:
raise ValueError('Token input hash mismatch: '+name)
with (a.audit/'documents.jsonl').open() as f:
audit_ids={json.loads(l)['id'] for l in f}
if len(audit_ids)!=6674:
raise ValueError('Expected exactly 6674 audited exclusion IDs')
a.output.mkdir(parents=True,exist_ok=False)
a.token_output.mkdir(parents=True,exist_ok=False)
manifest=dict(complete=False,started=now(),input=str(a.input.resolve()),
input_manifest_sha256=sha(a.input/'manifest.json'),
input_sha256=tm['input_sha256'],audit_report_sha256=sha(a.audit/'report.json'),
policy='Exclude exactly audited Kanripo files with unknown catalog dynasty; preserve all other rows verbatim',
script_sha256=sha(__file__))
write_json(a.output/'manifest.json',manifest)
token_manifest=dict(complete=False,started=now(),parent=str(a.tokens.resolve()),
parent_manifest_sha256=sha(a.tokens/'manifest.json'),
tokenizer_sha256=tm['tokenizer_sha256'],vocab_size=tm['vocab_size'],dtype=tm['dtype'],
script_sha256=sha(__file__),encoding=tm['encoding'],
policy='Copy retained document token slices exactly; no re-encoding')
write_json(a.token_output/'manifest.json',token_manifest)
original=np.memmap(a.tokens/'tokens.bin',dtype='<u4',mode='r')
kept,removed,token_counts,chars=Counter(),Counter(),Counter(),Counter()
removed_ids,removed_works=set(),set()
offset=0
parent_offset=0
with src.open('rb') as source, (a.tokens/'documents.jsonl').open() as metadata, \
(a.output/'pretrain.jsonl').open('wb') as corpus, \
(a.output/'excluded.jsonl').open('w') as exclusions, \
(a.token_output/'documents.jsonl').open('w') as documents, \
(a.token_output/'tokens.bin').open('wb') as tokens:
for raw in source:
r,m=json.loads(raw),json.loads(next(metadata))
if r['id']!=m['id'] or r['text_sha256']!=m['text_sha256'] or m['offset']!=parent_offset:
raise ValueError('Source/token metadata alignment failed')
parent_offset+=m['tokens']
if excluded(r)!=(r['id'] in audit_ids):
raise ValueError('Exclusion policy does not match audited IDs')
if excluded(r):
removed[r['source']]+=1
removed_ids.add(r['id']);removed_works.add(r['source_document_id'])
exclusions.write(json.dumps(dict(id=r['id'],source=r['source'],title=r['title'],
source_document_id=r['source_document_id'],text_sha256=r['text_sha256'],
tokens=m['tokens'],reason='user_excluded_kanripo_unknown_dynasty'),ensure_ascii=False)+'\n')
continue
if hashlib.sha256(r['text'].encode()).hexdigest()!=r['text_sha256']:
raise ValueError('Source document hash mismatch')
corpus.write(raw)
ids=original[m['offset']:m['offset']+m['tokens']]
if len(ids)!=m['tokens'] or ids[0]!=1 or ids[-1]!=2:
raise ValueError('Invalid retained tokens')
tokens.write(ids.tobytes())
m['offset']=offset
documents.write(json.dumps(m,ensure_ascii=False)+'\n')
kept[r['source']]+=1;token_counts[r['source']]+=m['tokens'];chars[r['source']]+=len(r['text'])
offset+=m['tokens']
if next(metadata,None) is not None or parent_offset!=len(original):
raise ValueError('Incomplete source coverage')
if removed_ids!=audit_ids or len(removed_works)!=1881:
raise ValueError('Excluded IDs/work count mismatch')
if sum(kept.values())!=tm['documents']-6674:
raise ValueError('Kept count mismatch')
for name in ('tokenizer.model','tokenizer_wrapper.py'):
shutil.copyfile(a.tokens/name,a.token_output/name)
# Independently compare every persisted slice and metadata row to its parent.
copied=np.memmap(a.token_output/'tokens.bin',dtype='<u4',mode='r')
verified=0
with (a.tokens/'documents.jsonl').open() as old, (a.token_output/'documents.jsonl').open() as new:
for line in old:
before=json.loads(line)
if before['id'] in audit_ids:
continue
after=json.loads(next(new))
old_offset,new_offset=before.pop('offset'),after.pop('offset')
if before!=after:
raise ValueError('Retained token metadata changed')
length=before['tokens']
if not np.array_equal(original[old_offset:old_offset+length],copied[new_offset:new_offset+length]):
raise ValueError('Retained token IDs changed')
verified+=1
if next(new,None) is not None:
raise ValueError('Unexpected retained rows')
# Re-read corpus and confirm exact retained source bytes and exclusion absence.
with src.open('rb') as old, (a.output/'pretrain.jsonl').open('rb') as new:
for line in old:
if json.loads(line)['id'] not in audit_ids and line!=next(new):
raise ValueError('Retained corpus bytes changed')
if next(new,None) is not None:
raise ValueError('Unexpected corpus rows')
verification=dict(complete=True,excluded_files=len(removed_ids),excluded_works=len(removed_works),
kept_files=verified,kept_tokens=offset,all_retained_text_bytes_identical=True,
all_retained_token_ids_identical=True,excluded_id_overlap=0)
write_json(a.output/'verification.json',verification)
write_json(a.token_output/'verification.json',verification)
report=dict(kept_units=sum(kept.values()),kept_characters=sum(chars.values()),tokens=offset,
excluded_files=len(removed_ids),excluded_works=len(removed_works),
removed_tokens=tm['tokens']-offset,
counts={s:dict(kept_units=n,kept_chars=chars[s],tokens=token_counts[s]) for s,n in kept.items()})
write_json(a.output/'report.json',report)
corpus_sha=sha(a.output/'pretrain.jsonl')
manifest.update(complete=True,finished=now(),output_sha256={p.name:sha(p) for p in a.output.iterdir() if p.name!='manifest.json'})
write_json(a.output/'manifest.json',manifest)
token_manifest.update(complete=True,finished=now(),input=str((a.output/'pretrain.jsonl').resolve()),
input_sha256=corpus_sha,documents=verified,tokens=offset,by_source=dict(token_counts),
files={p.name:sha(p) for p in a.token_output.iterdir() if p.name!='manifest.json'})
write_json(a.token_output/'manifest.json',token_manifest)
print(json.dumps(report,ensure_ascii=False,indent=2),flush=True)
if __name__=='__main__':
main()
|