Include reproducible source conversion

repackage_dataset/relevance.py

@@ -0,0 +1,199 @@
+"""Pinned entity sentiment and query relevance mirrors; no generated labels or negatives."""
+import ast
+from collections import Counter
+import csv
+import hashlib
+import heapq
+import io
+import json
+from pathlib import Path
+import re
+import urllib.request
+
+from datasets import Dataset, DatasetDict, Features, Value
+from huggingface_hub import snapshot_download
+import pyarrow.parquet as pq
+
+from .text import _split_of
+
+SENTFIN_REVISION = 'eba43610fe3cb57a3e5773cd6c037da1f8992ad5'
+WANDS_REVISION = '16cff4fbe4b46ee0b382e7901ee62461f88a2cb1'
+SEARCH_REVISION = '781d35d1bf87253b3dcd0fadcb82bfbee9c244f1'
+
+
+def text_key(text):
+    return hashlib.sha256(' '.join(text.casefold().split()).encode()).hexdigest()
+
+
+def parse_sentfin(value):
+    """Repair only the released unescaped-apostrophe dictionary representation."""
+    try:
+        result = ast.literal_eval(value)
+        repaired = False
+    except (ValueError, SyntaxError):
+        pattern = r"'(.*?)'\s*:\s*'(positive|neutral|negative)'(?=\s*[,}])"
+        pairs = re.findall(pattern, value)
+        remainder = re.sub(pattern, '', value).strip()
+        if not pairs or re.sub(r'[{},\s]', '', remainder) or len(dict(pairs)) != len(pairs):
+            raise ValueError('Unrecognized SEntFiN annotation representation')
+        result, repaired = dict(pairs), True
+    if not isinstance(result, dict) or not result or any(
+        not isinstance(k, str) or not k.strip() or v not in ('negative', 'neutral', 'positive')
+        for k, v in result.items()):
+        raise ValueError('Invalid SEntFiN entity sentiment')
+    return result, repaired
+
+
+def write_mirror(name, splits, label_type, report):
+    directory = Path('build') / (name + '-release')
+    directory.mkdir(parents=True, exist_ok=True)
+    features = Features(dict(sentence1=Value('string'), sentence2=Value('string'),
+                             labels=Value(label_type), metadata=Value('string')))
+    data = DatasetDict({split:Dataset.from_list(rows,features=features) for split,rows in splits.items()})
+    report.setdefault('conversion_sha256',hashlib.sha256(Path(__file__).read_bytes()).hexdigest())
+    report.update(splits={s:len(d) for s,d in data.items()})
+    for split, rows in data.items():
+        rows.to_parquet(str(directory / (split + '.parquet')))
+    (directory/'excluded-questions.jsonl').write_text('')
+    (directory/'provenance.json').write_text(json.dumps(report,indent=2,ensure_ascii=False)+'\n')
+    return data
+
+
+def sentfin():
+    """SEntFiN entity sentiment, one native headline/entity judgment per row.
+
+    Pinned original CSV: https://github.com/pyRis/SEntFiN (repository MIT license).
+    Only malformed dictionary quoting is repaired; the original annotation string
+    and source row ID remain in metadata. Identical normalized headlines stay in
+    one deterministic 80/10/10 internal partition. These are derived holdouts, not
+    an official benchmark split. Duplicate pairs are deduplicated; contradictory
+    judgments for the same headline/entity are excluded and reported.
+    """
+    url=f'https://raw.githubusercontent.com/pyRis/SEntFiN/{SENTFIN_REVISION}/SEntFiN.csv'
+    raw=urllib.request.urlopen(url).read()
+    records, repairs, conflicts, duplicates = {}, [], set(), 0
+    for row in csv.DictReader(io.StringIO(raw.decode())):
+        labels, repaired = parse_sentfin(row['Decisions'])
+        if repaired:
+            repairs.append(dict(id=row['S No.'], original=row['Decisions'], parsed=labels))
+        group=text_key(row['Title'])
+        for entity,label in labels.items():
+            key=(group,entity.casefold())
+            if key in records:
+                if records[key]['labels'] != label:
+                    conflicts.add(key)
+                else:
+                    duplicates += 1
+                continue
+            meta=dict(id=f"{row['S No.']}:{entity}", headline_id=row['S No.'], entity=entity,
+                split_group_id=group, source_url=url, source_revision=SENTFIN_REVISION,
+                source_annotation=row['Decisions'], quoting_repaired=repaired)
+            records[key]=dict(sentence1=row['Title'],sentence2=entity,labels=label,
+                              metadata=json.dumps(meta,ensure_ascii=False,sort_keys=True))
+    splits={s:[] for s in ('train','validation','test')}
+    for (group,entity),row in records.items():
+        if (group,entity) not in conflicts:
+            splits[_split_of(group)].append(row)
+    return write_mirror('sentfin',splits,'string',dict(source_url=url,revision=SENTFIN_REVISION,
+        source_sha256=hashlib.sha256(raw).hexdigest(),repairs=repairs,duplicate_pairs=duplicates,
+        excluded_conflicting_pairs=[list(k) for k in sorted(conflicts)],
+        split_policy='normalized-headline-hash-80/10/10',license='mit',
+        license_url=f'https://github.com/pyRis/SEntFiN/blob/{SENTFIN_REVISION}/LICENSE'))
+
+
+def wands():
+    """WANDS native three-grade product relevance with query-disjoint internal holdouts.
+
+    Original Wayfair data: https://github.com/wayfair/WANDS (MIT).
+    Uses the pinned joined napsternxg/wands Hub copy; its random row splits overlap
+    on queries and are replaced by deterministic 80/10/10 normalized-query groups.
+    All native judgments are retained. Inputs contain query and product content;
+    popularity/rating fields and relevance labels never enter model inputs.
+    """
+    root=Path(snapshot_download('napsternxg/wands',repo_type='dataset',revision=WANDS_REVISION,
+                               allow_patterns=['data/*.parquet']))
+    splits={s:[] for s in ('train','validation','test')}
+    counts=Counter()
+    for path in sorted(root.glob('data/*.parquet')):
+        for batch in pq.ParquetFile(path).iter_batches(batch_size=512):
+            for row in batch.to_pylist():
+                if row['label'] not in (0,1,2):
+                    raise ValueError('Unknown WANDS relevance grade')
+                group=text_key(row['query'])
+                fields=('product_name','product_class','category hierarchy','product_description','product_features')
+                product='\n'.join(f'{key}: {row[key]}' for key in fields if row.get(key))
+                meta=dict(id=str(row['id']),query_id=row['query_id'],product_id=row['product_id'],
+                    split_group_id=group,original_mirror_split=path.name.split('-')[0],
+                    source_dataset='napsternxg/wands',source_revision=WANDS_REVISION)
+                splits[_split_of(group)].append(dict(sentence1=row['query'],sentence2=product,
+                    labels=row['label'],metadata=json.dumps(meta,sort_keys=True)))
+                counts[row['label']] += 1
+    return write_mirror('wands',splits,'int64',dict(source='napsternxg/wands',revision=WANDS_REVISION,
+        original_source='https://github.com/wayfair/WANDS',label_counts=dict(counts),
+        source_sha256={p.name:hashlib.sha256(p.read_bytes()).hexdigest() for p in root.glob('data/*.parquet')},
+        split_policy='normalized-query-hash-80/10/10',license='mit',
+        license_url='https://github.com/wayfair/WANDS/blob/3b74dcf4ba29ab8ff3e6a50b5b09fc627cb882b5/LICENSE'))
+
+
+def search_record(row, candidate):
+    score=candidate['score']
+    if not isinstance(score,(int,float)) or not 0 <= score <= 14:
+        raise ValueError(f'Unexpected SciRepEval click-derived score: {score}')
+    document='\n'.join(v for v in (candidate.get('title'),candidate.get('abstract')) if v)
+    if not document.strip() or not row['query'].strip():
+        return None
+    group=text_key(row['query'])
+    meta=dict(id=f"{row['doc_id']}:{candidate['doc_id']}",query_id=row['doc_id'],
+        candidate_id=candidate['doc_id'],split_group_id=group,source_revision=SEARCH_REVISION,
+        source_score=score,supervision='click-derived score, not a human relevance rating',
+        score_anchors=list(range(0,15,2)),score_interpolation='linear; expected score equals native value')
+    return dict(sentence1=row['query'],sentence2=document,labels=float(score),metadata=json.dumps(meta,sort_keys=True))
+
+
+def scirepeval_search(max_rows=5000, max_rows_eval=500, seed=42):
+    """SciRepEval Search native train/validation pairs with raw click-derived numeric scores.
+
+    No evaluation/test benchmark data is consumed. All training/validation shards
+    are scanned; deterministic hash sampling covers eligible pairs from the entire
+    source, not an early prefix. Validation queries seen in training are excluded.
+    Scores remain unchanged. Jev uses linear interpolation over eight anchors
+    0,2,...,14 to preserve the numeric expectation within its 10-level limit.
+    Dataset/text licensing is unspecified: the code's Apache license is not used
+    as a grant for paper abstracts. Evidence and exclusions remain in provenance.
+    """
+    root=Path(snapshot_download('allenai/scirepeval',repo_type='dataset',revision=SEARCH_REVISION,
+        allow_patterns=['search/train-*.parquet','search/validation-*.parquet'],max_workers=4))
+    splits, counts, train_groups = {}, Counter(), set()
+    scores=Counter()
+    for split,limit in [('train',max_rows),('validation',max_rows_eval)]:
+        pool=[]
+        for path in sorted(root.glob(f'search/{split}-*.parquet')):
+            for batch in pq.ParquetFile(path).iter_batches(batch_size=64):
+                for source in batch.to_pylist():
+                    group=text_key(source['query'])
+                    if split=='validation' and group in train_groups:
+                        counts['validation/excluded/query_overlap'] += len(source['candidates'])
+                        continue
+                    if split=='train':
+                        train_groups.add(group)
+                    for candidate in source['candidates']:
+                        record=search_record(source,candidate)
+                        if record is None:
+                            counts[split+'/excluded/missing_text'] += 1
+                            continue
+                        counts[split+'/eligible'] += 1
+                        scores[candidate['score']] += 1
+                        identity=f"{source['doc_id']}:{candidate['doc_id']}"
+                        rank=-int(hashlib.sha256(f'{seed}:{identity}'.encode()).hexdigest(),16)
+                        item=(rank,identity,counts[split+'/eligible'],record)
+                        if len(pool)<limit:
+                            heapq.heappush(pool,item)
+                        elif item[:2]>pool[0][:2]:
+                            heapq.heapreplace(pool,item)
+        # Duplicate native query/candidate pairs must not become duplicated training examples.
+        selected={identity:record for _,identity,_,record in sorted(pool)}
+        splits[split]=list(selected.values())
+    return write_mirror('scirepeval-search',splits,'float64',dict(source='allenai/scirepeval',
+        revision=SEARCH_REVISION,config='search',counts=dict(counts),source_score_counts=dict(scores),
+        selection=dict(max_rows=max_rows,max_rows_eval=max_rows_eval,seed=seed,method='full-source hash sample'),
+        split_policy='native train/validation; validation queries in train excluded',license='unspecified'))