Datasets:
File size: 3,402 Bytes
d9a8780 | 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 | """MIT: read an explicitly selected canonical split from original text shards.
No training on the all-documents view by accident: split is mandatory.
"""
from itertools import zip_longest
import hashlib
import json
from pathlib import Path
import pyarrow.parquet as pq
def document_id(inventory_id,source,original_file,row):
encoded=json.dumps(['release-doc-id-v1',inventory_id,source,original_file,row],
ensure_ascii=False,separators=(',',':')).encode('utf8')
return hashlib.sha256(encoded).hexdigest()
def iter_split(text_path,assignment_path,*,split,source,original_file,inventory_id,batch_size=8192):
if split=='validation':split='eval'
if split not in ('train','eval'):raise ValueError('Explicit train or validation split required')
text=pq.ParquetFile(text_path);assignment=pq.ParquetFile(assignment_path)
if text.metadata.num_rows!=assignment.metadata.num_rows:raise ValueError('Text/assignment coverage mismatch')
metadata=assignment.schema_arrow.metadata or {}
if metadata.get(b'original_final_file')!=original_file.encode():raise ValueError('Assignment belongs to another shard')
row_number=0
for tb,ab in zip_longest(text.iter_batches(batch_size=batch_size,use_threads=False),
assignment.iter_batches(batch_size=batch_size,use_threads=False)):
if tb is None or ab is None or tb.num_rows!=ab.num_rows:raise ValueError('Batch coverage mismatch')
for row,assigned in zip(tb.to_pylist(),ab.to_pylist()):
if assigned['row']!=row_number:raise ValueError('Assignment row order mismatch')
if hashlib.sha256(row['text'].encode('utf8')).digest()!=assigned['content_sha256']:
raise ValueError('Document body differs from canonical split')
if assigned['split']==split:
yield {**row,'doc_id':document_id(inventory_id,source,original_file,row_number),
'source':source,'original_row':row_number,'split':split,'group_id':assigned['group_id']}
elif assigned['split'] not in ('train','eval'):raise ValueError('Unknown canonical split')
row_number+=1
def download_and_iter(repo_id,*,revision,split,max_documents=None):
"""Authenticated downloads are unnecessary for the public release."""
from huggingface_hub import hf_hub_download
def fetch(path):return hf_hub_download(repo_id,path,repo_type='dataset',revision=revision,token=False)
manifest=json.loads(Path(fetch('manifest.json')).read_text());done=0
if not manifest.get('source_documents'):raise ValueError('No text shards listed')
for item in manifest['source_documents']:
tp=fetch(item['text_path']);ap=fetch(item['assignment_path'])
# Bind cached files to release bytes before decoding.
for p,wanted in [(tp,item['text_sha256']),(ap,item['assignment_sha256'])]:
h=hashlib.sha256()
with open(p,'rb') as f:
while block:=f.read(8*1024**2):h.update(block)
if h.hexdigest()!=wanted:raise ValueError('Downloaded file SHA mismatch')
for row in iter_split(tp,ap,split=split,source=manifest['source'],original_file=item['original_final_file'],
inventory_id=manifest['content_inventory_id']):
yield row;done+=1
if max_documents is not None and done>=max_documents:return
|