scripts/convert_dataset.py

"""Stream original JSON into image-embedded QA Parquet, many dynamically scheduled record chunks per source."""
import argparse
import ast
from collections import Counter
from concurrent.futures import ProcessPoolExecutor, as_completed
import hashlib
import io
import json
import logging
import multiprocessing
import os
from pathlib import Path
import re
import shutil
import sys

import ijson
import pyarrow as pa
import pyarrow.parquet as pq
from datasets import Features, Image, Value
from filelock import FileLock
from PIL import Image as PILImage
import yaml

VERSION = 2
FIELDS = Features({'image': Image(), 'question': Value('string'), 'answer': Value('string'),
                   'type': Value('string'), 'source': Value('string'), 'task': Value('string'),
                   'id': Value('string'), 'qa_index': Value('int32'), 'record_index': Value('int64')})
SCHEMA = FIELDS.arrow_schema


def sha256(path):
    h = hashlib.sha256()
    with Path(path).open('rb') as f:
        for block in iter(lambda: f.read(8 * 1024 * 1024), b''):
            h.update(block)
    return h.hexdigest()


def save_json(path, value):
    path = Path(path)
    tmp = path.with_suffix(path.suffix + '.tmp')
    tmp.write_text(json.dumps(value, ensure_ascii=False, indent=2), encoding='utf-8')
    os.replace(tmp, path)


def infer_task(question, answer):
    """Conservative derived label; never rewrite or execute answer text."""
    if len(answer) > 100000:
        return None
    try:
        points = ast.literal_eval(answer)
        if not isinstance(points, (list, tuple)) or not points:
            return None
        if any(not isinstance(p, (list, tuple)) or len(p) not in (2, 3) for p in points):
            return None
        dims = {len(p) for p in points}
        if len(dims) != 1 or any(not isinstance(v, (int, float)) or isinstance(v, bool) for p in points for v in p):
            return None
    except (ValueError, SyntaxError, TypeError, RecursionError):
        return None
    q = re.sub(r'\b([23])\s+d\b', r'\1d', question.lower())
    if re.search(r'\blift(?:ing)?\b', q) and '2d' in q and '3d' in q and dims == {3}:
        return '2d_to_3d'
    if dims == {2} and '2d' in q:
        return '2d_trace'
    if dims == {3} and '3d' in q and '2d' not in q:
        return '3d_trace'
    return None


def resolve_image(root, name):
    # Input release uses flat image names. Do not accept traversal or absolute paths.
    if not isinstance(name, str) or '/' in name or '\\' in name or name in ('', '.', '..'):
        raise ValueError(f'Expected flat image filename: {name!r}')
    candidates = [name] if ':' not in name else [name.replace(':', '_')]
    for candidate in candidates:
        path = (root / candidate).resolve()
        if path.parent == root and path.is_file():
            return path
    raise FileNotFoundError(f'RGB image missing: {name}')


def convert_source(job):
    source, metadata, images, out, opts, chunk_id, start_index = job
    metadata, images, out = Path(metadata), Path(images), Path(out)
    pa.set_cpu_count(1)  # prevent each process spawning another CPU-sized thread pool
    dest = out / 'data' / source
    dest.mkdir(parents=True, exist_ok=True)
    state_path = out / '.state' / f'{source}-{chunk_id:05d}.json'
    stat = metadata.stat()
    identity = {'version': VERSION, 'script_sha256': sha256(__file__), 'metadata': str(metadata),
                'size': stat.st_size, 'mtime_ns': stat.st_mtime_ns,
                'images': str(images), 'options': opts, 'chunk_id': chunk_id, 'start_index': start_index}
    if state_path.exists():
        state = json.loads(state_path.read_text(encoding='utf-8'))
        if state['identity'] != identity:
            raise RuntimeError(f'{source}: input/options/code changed; use a new output directory')
        for item in state['files']:
            path = out / item['path']
            if not path.is_file() or path.stat().st_size != item['bytes'] or sha256(path) != item['sha256']:
                raise RuntimeError(f'{source}: completed shard changed: {path}')
        if state['complete']:
            print(f'{source}: resume verified, already complete', flush=True)
            return state
    else:
        state = {'identity': identity, 'source': source, 'complete': False, 'next_record': 0,
                 'worker_pid': os.getpid(), 'chunk_id': chunk_id, 'qa_rows': 0, 'mapped_image_records': 0, 'task_counts': {}, 'type_counts': {}, 'files': []}
        save_json(state_path, state)
    writer = None
    pending, pending_size, shard_size, shard_rows = [], 0, 0, 0
    tasks, types = Counter(state['task_counts']), Counter(state['type_counts'])
    processed, total_rows = state['next_record'], state['qa_rows']
    mapped = state['mapped_image_records']
    target = opts['shard_mib'] * 1024**2
    group_target = opts['row_group_mib'] * 1024**2
    shard_path = None
    tmp = None

    def flush_group():
        nonlocal pending, pending_size, writer, shard_path, tmp
        if not pending:
            return
        if writer is None:
            shard_path = dest / f'train-{chunk_id:05d}-{len(state["files"]):03d}.parquet'
            tmp = shard_path.with_suffix('.parquet.incomplete')
            writer = pq.ParquetWriter(tmp, SCHEMA, compression='zstd', compression_level=3,
                                     use_dictionary=True, dictionary_pagesize_limit=16*1024**2)
        writer.write_table(pa.Table.from_pylist(pending, schema=SCHEMA), row_group_size=len(pending))
        pending, pending_size = [], 0

    def finish_shard():
        nonlocal writer, shard_rows, shard_size
        flush_group()
        if writer is None:
            return
        writer.close()
        writer = None
        assert pq.ParquetFile(tmp).metadata.num_rows == shard_rows
        os.replace(tmp, shard_path)
        state['files'].append({'path': shard_path.relative_to(out).as_posix(),
                               'rows': shard_rows, 'bytes': shard_path.stat().st_size,
                               'sha256': sha256(shard_path)})
        state.update(next_record=processed, qa_rows=total_rows, mapped_image_records=mapped,
                     task_counts=dict(tasks), type_counts=dict(types))
        save_json(state_path, state)
        print(f'{source}: {processed:,} records / {total_rows:,} QA; committed {shard_path.name}', flush=True)
        shard_rows, shard_size = 0, 0

    try:
        with metadata.open('r', encoding='utf-8') as fp:
            for local_index, line in enumerate(fp):
                if local_index < state['next_record']:
                    continue
                record = json.loads(line)
                record_index = start_index + local_index
                refs = record.get('image')
                conv = record.get('conversations')
                if not isinstance(refs, list) or len(refs) != 1:
                    raise ValueError(f'{source}[{record_index}]: expected exactly one RGB reference')
                if not isinstance(conv, list) or not conv or len(conv) % 2:
                    raise ValueError(f'{source}[{record_index}]: invalid conversation length')
                path = resolve_image(images, refs[0])
                blob = path.read_bytes()
                if len(blob) > 64 * 1024**2:
                    raise ValueError(f'Image exceeds 64 MiB memory guard: {path}')
                if opts['verify_images']:
                    with PILImage.open(io.BytesIO(blob)) as im:
                        im.verify()
                mapped += path.name != refs[0]
                kind = record.get('type')
                if kind is not None and not isinstance(kind, str):
                    raise ValueError(f'{source}[{record_index}]: non-string type')
                if not isinstance(record.get('id'), str):
                    raise ValueError(f'{source}[{record_index}]: non-string id')
                types[kind if kind is not None else '(missing)'] += 1
                for offset in range(0, len(conv), 2):
                    q, a = conv[offset:offset+2]
                    if q.get('from') != 'human' or a.get('from') != 'gpt':
                        raise ValueError(f'{source}[{record_index}]: nonalternating roles')
                    question, answer = q['value'], a['value']
                    if not isinstance(question, str) or not isinstance(answer, str):
                        raise ValueError('QA text must be strings')
                    task = infer_task(question, answer)
                    tasks[task or '(unknown)'] += 1
                    row = {'image': {'bytes': blob, 'path': path.name}, 'question': question,
                           'answer': answer, 'type': kind, 'source': source, 'task': task,
                           'id': record['id'], 'qa_index': offset//2, 'record_index': record_index}
                    size = len(blob) + len(question.encode('utf-8')) + len(answer.encode('utf-8')) + 256
                    if pending and (pending_size + size > group_target or len(pending) >= 128):
                        flush_group()
                    pending.append(row)
                    pending_size += size
                    shard_size += size
                    shard_rows += 1
                    total_rows += 1
                processed = local_index + 1
                # Checkpoint only at whole-record boundaries, preserving conversation groups.
                if shard_size >= target:
                    finish_shard()
        finish_shard()
        state.update(complete=True, next_record=processed, qa_rows=total_rows,
                     mapped_image_records=mapped, task_counts=dict(tasks), type_counts=dict(types))
        save_json(state_path, state)
        return state
    except BaseException:
        if writer is not None:
            writer.close()
        raise


def write_card(out, manifest):
    template = Path(__file__).with_name('README.template.md').read_text(encoding='utf-8')
    configs = []
    if manifest['limited']:
        configs.append({'config_name': 'preview', 'default': True,
                        'data_files': [{'split': 'train', 'path': 'data/*/*.parquet'}]})
    else:
        configs.append({'config_name': 'all', 'default': True,
                        'data_files': [{'split': 'train', 'path': 'data/*/*.parquet'}]})
    for source in manifest['sources']:
        configs.append({'config_name': source, 'data_files': [{'split': 'train', 'path': f'data/{source}/*.parquet'}]})
    metadata = {'language': ['en'], 'pretty_name': 'TraceSpatial-Trace',
                'task_categories': ['visual-question-answering'],
                'tags': ['robotics', 'spatial-reasoning', 'trajectory', 'arxiv:2512.13660'], 'configs': configs}
    if manifest['limited']:
        status = f"This is a **preview** containing {manifest['qa_rows']:,} QA rows from {manifest['records']:,} original records. Full local-corpus counts are shown below; those are not the size of this preview."
    else:
        status = f"This conversion contains **{manifest['qa_rows']:,} QA rows** from **{manifest['records']:,} original records** across {len(manifest['sources'])} sources."
    text = template.replace('@@STATUS@@', status).replace('@@CONFIG@@', 'preview' if manifest['limited'] else 'all')
    (out/'README.md').write_text('---\n'+yaml.safe_dump(metadata, sort_keys=False)+'---\n\n'+text, encoding='utf-8')


def prepare_chunks(files, sources, out, options):
    """One streaming pass creates immutable, reusable work chunks (no RGB copy)."""
    index_path = out/'.state'/'chunks'/'index.json'
    if index_path.exists():
        chunks = json.loads(index_path.read_text(encoding='utf-8'))
        for c in chunks:
            path = Path(c['path'])
            if not path.is_file() or sha256(path) != c['sha256']:
                raise RuntimeError(f'Prepared JSONL chunk changed: {path}')
        return chunks
    chunks = []
    for source, metadata in zip(sources, files):
        directory = index_path.parent/source
        directory.mkdir(parents=True, exist_ok=True)
        batch, start, chunk_id = [], 0, 0
        def emit():
            nonlocal batch, start, chunk_id
            path = directory/f'chunk-{chunk_id:05d}.jsonl'
            tmp = path.with_suffix('.jsonl.tmp')
            with tmp.open('w', encoding='utf-8', newline='\n') as fp:
                for record in batch:
                    fp.write(json.dumps(record, ensure_ascii=False)+'\n')
            os.replace(tmp, path)
            chunks.append({'source':source, 'path':str(path), 'chunk_id':chunk_id,
                           'start_index':start, 'records':len(batch), 'sha256':sha256(path)})
            start += len(batch)
            chunk_id += 1
            batch = []
        with metadata.open('rb') as fp:
            for index, row in enumerate(ijson.items(fp, 'item')):
                if options['limit_records'] and index >= options['limit_records']:
                    break
                batch.append({key:row.get(key) for key in ('id','image','conversations','type')})
                if len(batch) >= options['chunk_records']:
                    emit()
                if (index+1) % 50000 == 0:
                    logging.info('%s: prepared %s original records',source,index+1)
        if batch: emit()
        if start == 0: raise ValueError(f'Empty metadata: {metadata}')
        logging.info('%s: prepared %s chunks / %s records',source,chunk_id,start)
    save_json(index_path,chunks)
    return chunks


def main():
    p = argparse.ArgumentParser(description=__doc__)
    p.add_argument('--input', type=Path, required=True, help='Directory containing images/ and metadata/')
    p.add_argument('--output', type=Path, required=True, help='New release directory; repeat to resume')
    p.add_argument('--workers', type=int, default=min(24, os.cpu_count() or 1))
    p.add_argument('--chunk-records', type=int, default=2048, help='Original records per independent work chunk; shared across all workers')
    p.add_argument('--shard-mib', type=int, default=512, help='Approximate uncompressed bytes; finish whole record before splitting')
    p.add_argument('--row-group-mib', type=int, default=32)
    p.add_argument('--limit-records', type=int, default=0, help='Per-source record limit; 0 means all')
    p.add_argument('--verify-images', action='store_true', help='Decode/verify every referenced image; slower')
    args = p.parse_args()
    if min(args.workers, args.shard_mib, args.row_group_mib, args.chunk_records) < 1 or args.limit_records < 0:
        p.error('worker/size arguments must be positive and limit nonnegative')
    root, out = args.input.resolve(), args.output.resolve()
    if out == root or root in out.parents:
        p.error('Output must be separate from the input dataset')
    files = sorted((root/'metadata').glob('*.json'))
    if not files or not (root/'images').is_dir():
        p.error('Input needs metadata/*.json and images/')
    sources = [f.name.split('_')[0] for f in files]
    if len(sources) != len(set(sources)):
        p.error('Multiple metadata files share a source label')
    out.mkdir(parents=True, exist_ok=True)
    (out/'.state').mkdir(exist_ok=True)
    logging.basicConfig(level=logging.INFO, format='%(asctime)s %(message)s',
                        handlers=[logging.StreamHandler(), logging.FileHandler(out/'.state'/'conversion.log', encoding='utf-8')])
    opts = {k: getattr(args, k) for k in ('shard_mib','row_group_mib','limit_records','verify_images','chunk_records')}
    with FileLock(out/'.state'/'convert.lock', timeout=0):
        plan = {'sources': sources, 'input': str(root), 'options': opts, 'script_sha256':sha256(__file__), 'metadata':[{ 'path':str(f), 'size':f.stat().st_size, 'mtime_ns':f.stat().st_mtime_ns} for f in files]}
        plan_path = out/'.state'/'plan.json'
        if plan_path.exists() and json.loads(plan_path.read_text(encoding='utf-8')) != plan:
            raise RuntimeError('Output belongs to another conversion; use a fresh directory')
        save_json(plan_path, plan)
        chunks = prepare_chunks(files, sources, out, opts)
        jobs = [(c['source'],c['path'],str((root/'images').resolve()),str(out),opts,c['chunk_id'],c['start_index']) for c in chunks]
        logging.info('Starting %s conversion processes for %s independent chunks',min(args.workers,len(jobs)),len(jobs))
        results = {}
        with ProcessPoolExecutor(max_workers=min(args.workers,len(jobs)), mp_context=multiprocessing.get_context('spawn')) as pool:
            futures = {pool.submit(convert_source,j):(j[0],j[5]) for j in jobs}
            for future in as_completed(futures):
                key = futures[future]
                try:
                    results[key] = future.result()
                    logging.info('%s chunk %s complete: %s QA rows (%s/%s chunks)',key[0],key[1],results[key]['qa_rows'],len(results),len(jobs))
                except BaseException:
                    for f in futures: f.cancel()
                    logging.exception('%s failed; completed shards remain resumable',key)
                    raise
        ordered = [results[key] for key in sorted(results)]
        aggregate = {}
        for source in sources:
            group = [r for r in ordered if r['source']==source]
            summary = {k:sum(r[k] for r in group) for k in ('next_record','qa_rows','mapped_image_records')}
            for key in ('task_counts','type_counts'):
                counter=Counter()
                for r in group: counter.update(r[key])
                summary[key]=dict(counter)
            aggregate[source]=summary
        manifest = {'version':VERSION, 'complete':True, 'limited':bool(args.limit_records),
                    'sources':sources, 'records':sum(r['next_record'] for r in ordered),
                    'qa_rows':sum(r['qa_rows'] for r in ordered),
                    'files':[item for r in ordered for item in r['files']],
                    'source_statistics':aggregate,
                    'conversion':{'requested_workers':args.workers, 'chunks':len(jobs),
                                  'worker_pids':sorted({r['worker_pid'] for r in ordered})}}
        write_card(out,manifest)
        script_dest = out/'scripts'
        script_dest.mkdir(exist_ok=True)
        for f in Path(__file__).parent.iterdir():
            if f.is_file() and f.suffix in ('.py','.md','.txt','.yml') and f.resolve() != (script_dest/f.name).resolve():
                shutil.copy2(f,script_dest/f.name)
        save_json(out/'manifest.json', manifest)  # publish only after all sources succeed
        logging.info('COMPLETE: %s QA rows, %s shards',manifest['qa_rows'],len(manifest['files']))


if __name__ == '__main__':
    multiprocessing.freeze_support()
    main()