crosscodeeval_v2/patch.py

"""Reproduce CrossCodeEval v2 tasks.parquet: the current v1 payload of every kept task plus the recorded results.

pip install pyarrow huggingface_hub
python patch.py --out ./rebuilt

Both inputs are pinned to repository revision a43f5635 and checked by SHA-256:
- crosscodeeval_v1/tasks.parquet: the current v1 payloads, content hashes and image manifest.
- crosscodeeval_v2/tasks.parquet as published before this refresh: which tasks are kept, the
  recorded success counts (`instruct_successes`, `thinking_successes`, `attempts_per_model`, `archive_category`) and the release metadata.
Use --source-dir PATH (holding v1.parquet and v2.parquet) to reuse local copies.
Trajectories and success counts stay as recorded; each trace row keeps the content hash it
was evaluated on.
"""
import argparse
import copy
import hashlib
import json
from pathlib import Path

import pyarrow as pa
import pyarrow.parquet as pq

REPO = 'FWeindel/validated-tasks'
REVISION = 'a43f56353751d6d73a93ff448de6981040c4d0c5'
SOURCES = {'v1': ('crosscodeeval_v1/tasks.parquet', '2607baee26489948dbfeb1a02494e5a7d149752a1910be882c40cbc09f139f0f'),
           'v2': ('crosscodeeval_v2/tasks.parquet', '21a25eac1b374c251f67f76e7abbb865435eeba329c7598ba210ba274d31f648')}
REFRESHED = '2026-10-17'


def source(name, source_dir, cache):
    path_in_repo, digest = SOURCES[name]
    if source_dir is None:
        from huggingface_hub import hf_hub_download
        path = Path(hf_hub_download(REPO, path_in_repo, repo_type='dataset', revision=REVISION, cache_dir=cache))
    else:
        path = source_dir / f'{name}.parquet'
    if hashlib.sha256(path.read_bytes()).hexdigest() != digest:
        raise ValueError(f'pinned source hash mismatch: {path_in_repo}')
    return path


def restrict_images(manifest, kept):
    result = copy.deepcopy(manifest)
    result['tasks'] = {path: entry for path, entry in manifest['tasks'].items() if path in kept}
    used = {env['bundle'] for entry in result['tasks'].values() for env in entry['environments']}
    result['bundles'] = {bid: bundle for bid, bundle in manifest['bundles'].items() if bid in used}
    result['missing'] = [m for m in manifest.get('missing', []) if m.get('task') in kept]
    return result


def build(source_dir, out):
    out.mkdir(parents=True, exist_ok=True)
    if (out / 'tasks.parquet').exists():
        raise ValueError('tasks.parquet already exists; choose an empty output directory')
    v1 = pq.read_table(source('v1', source_dir, out / '.source-cache'))
    old = pq.read_table(source('v2', source_dir, out / '.source-cache'))
    current = {row['path']: row for row in v1.to_pylist()}
    extra = [name for name in old.schema.names if name not in v1.schema.names]
    rows, earlier = [], 0
    for row in old.to_pylist():
        fresh = dict(current[row['path']])
        fresh.update({name: row[name] for name in extra})
        # The published v2 hash is the one the teacher runs were scored on.
        earlier += fresh['content_sha256'] != row['content_sha256']
        rows.append(fresh)
    schema = pa.schema([*v1.schema, *(old.schema.field(name) for name in extra)])
    metadata = dict(old.schema.metadata or {})
    if b'ot.images.v1' in (v1.schema.metadata or {}):
        images = restrict_images(json.loads(v1.schema.metadata[b'ot.images.v1']), {r['path'] for r in rows})
        metadata[b'ot.images.v1'] = json.dumps(images).encode()
    release = json.loads(metadata.get(b'ot.release.v2', b'{}'))
    release['current_v1'] = {'folder': 'crosscodeeval_v1', 'revision': REVISION, 'sha256': SOURCES['v1'][1],
                             'refreshed': REFRESHED, 'kept_tasks_evaluated_on_earlier_version': earlier}
    metadata[b'ot.release.v2'] = json.dumps(release).encode()
    pq.write_table(pa.Table.from_pylist(rows, schema=schema).replace_schema_metadata(metadata),
                   out / 'tasks.parquet', compression='zstd')
    print(json.dumps({'kept': len(rows), 'evaluated_on_earlier_version': earlier,
                      'images': b'ot.images.v1' in metadata}))


if __name__ == '__main__':
    parser = argparse.ArgumentParser(description=__doc__, formatter_class=argparse.RawDescriptionHelpFormatter)
    parser.add_argument('--source-dir', type=Path)
    parser.add_argument('--out', type=Path, required=True)
    args = parser.parse_args()
    build(args.source_dir, args.out)