reconstruction/reconstruct_coedit.py

"""Reconstruct CES CoEdIT inputs locally from the official upstream distribution.

Requires huggingface_hub and pyarrow. Downloading requires an explicit flag;
users must establish that their intended use complies with upstream terms.
The output directory must be separate from the public artifact tree.
"""

import argparse
import hashlib
import json
from pathlib import Path


def sha256(data):
    return hashlib.sha256(data).hexdigest()


def read_jsonl(path):
    with path.open(encoding='utf-8') as stream:
        return [json.loads(line) for line in stream if line.strip()]


def strip_instruction(text):
    prefix, separator, rest = text.partition(': ')
    return rest if separator and len(prefix.split()) <= 8 else text


def load_sources(folder, specification):
    sources = {}
    for split, expected in specification['upstream_files'].items():
        path = folder / f'{split}.jsonl'
        if sha256(path.read_bytes()) != expected['sha256']:
            raise ValueError(f'Upstream file hash mismatch: {path}')
        for row in read_jsonl(path):
            key = (split, str(row['_id']))
            if key in sources:
                raise ValueError(f'Duplicate upstream ID: {key}')
            sources[key] = (strip_instruction(row['src']), row['tgt'])
    return sources


def source_for(reference, sources):
    pair = sources[(reference['split'], reference['id'])]
    text = pair[reference['field']]
    if sha256(text.encode()) != reference['sha256']:
        raise ValueError(f'Source text mismatch: {reference["split"]}/{reference["id"]}')
    return text


def materialize(parts, source):
    chunks = []
    for part in parts:
        if set(part) == {'slice'}:
            start, end = part['slice']
            if not 0 <= start <= end <= len(source):
                raise ValueError('Invalid source slice')
            chunks.append(source[start:end])
        elif set(part) == {'literal'}:
            chunks.append(part['literal'])
        else:
            raise ValueError('Unknown edit payload format')
    return ''.join(chunks)


def write_jsonl(path, rows):
    path.parent.mkdir(parents=True, exist_ok=True)
    with path.open('w', encoding='utf-8') as stream:
        for row in rows:
            stream.write(json.dumps(row, ensure_ascii=False, separators=(',', ':')) + '\n')


def reconstruct(recipe_dir, upstream_dir, output):
    import pyarrow as pa
    import pyarrow.parquet as pq

    spec = json.loads((recipe_dir / 'specification.json').read_text())
    sources = load_sources(upstream_dir, spec)
    counts = {}
    for item in spec['manifests']:
        rows = []
        for recipe in read_jsonl(recipe_dir / item['recipe']):
            row = dict(recipe['metadata'])
            for field, reference in recipe['text_references'].items():
                row[field] = source_for(reference, sources)
            rows.append(row)
        destination = output / item['output']
        write_jsonl(destination, rows)
        counts[item['output']] = len(rows)
    for item in spec['clean_shards']:
        rows = []
        for recipe in read_jsonl(recipe_dir / item['recipe']):
            row = dict(recipe['metadata'])
            for field, reference in recipe['text_references'].items():
                row[field] = source_for(reference, sources)
            rows.append(row)
        destination = output / item['output']
        destination.parent.mkdir(parents=True, exist_ok=True)
        pq.write_table(pa.Table.from_pylist(rows), destination, compression='zstd')
        counts[item['output']] = len(rows)
    rows = []
    for recipe in read_jsonl(recipe_dir / 'corruptions.jsonl'):
        source = source_for(recipe['source'], sources)
        corrupted = materialize(recipe['corrupted_parts'], source)
        if sha256(corrupted.encode()) != recipe['corrupted_sha256']:
            raise ValueError('Reconstructed corruption hash mismatch')
        row = dict(recipe['metadata'])
        row['source_text'] = source
        row['corrupted_text'] = corrupted
        for direction, text in [('forward', source), ('inverse', corrupted)]:
            edits = []
            for encoded in recipe[f'{direction}_edits']:
                edit = dict(encoded)
                if 'content_parts' in edit:
                    edit['content'] = materialize(edit.pop('content_parts'), text)
                edits.append(edit)
            row[f'{direction}_edits_json'] = json.dumps(edits)
        rows.append(row)
    relative = 'data/posttraining/ces/v1/gec-corruptions/approved-corruptions.parquet'
    destination = output / relative
    destination.parent.mkdir(parents=True, exist_ok=True)
    pq.write_table(pa.Table.from_pylist(rows), destination, compression='zstd')
    counts[relative] = len(rows)
    (output / 'RECONSTRUCTION.json').write_text(json.dumps({
        'upstream_revision': spec['upstream_revision'], 'rows': counts,
        'scope': 'Text, row order and training edit programs verified; Parquet bytes and API traces are not reproduced.',
    }, indent=2) + '\n')
    return counts


def main():
    parser = argparse.ArgumentParser(description=__doc__)
    parser.add_argument('--upstream-dir', type=Path, required=True)
    parser.add_argument('--output', type=Path, required=True)
    parser.add_argument('--download', action='store_true', help='Fetch the pinned official CoEdIT files')
    args = parser.parse_args()
    recipes = Path(__file__).resolve().parent
    output = args.output.resolve()
    public_root = recipes.parent
    if output == public_root or public_root in output.parents or output in public_root.parents:
        parser.error('Reconstruct outside the public artifact tree to avoid accidental re-upload.')
    if output.exists() and any(output.iterdir()):
        parser.error('Choose an empty output directory; existing inputs will not be overwritten.')
    spec = json.loads((recipes / 'specification.json').read_text())
    upstream = args.upstream_dir.resolve()
    if upstream == public_root or public_root in upstream.parents:
        parser.error('Keep downloaded upstream files outside the public artifact tree.')
    if args.download:
        from huggingface_hub import hf_hub_download
        for filename in ('train.jsonl', 'validation.jsonl'):
            hf_hub_download(spec['upstream_repository'], filename,
                repo_type='dataset', revision=spec['upstream_revision'], local_dir=args.upstream_dir)
    counts = reconstruct(recipes, args.upstream_dir, output)
    print(json.dumps(counts, indent=2))


if __name__ == '__main__':
    main()