| |
| """Run the released SEGALE-COMET pipeline on the example or benchmark data.""" |
| import argparse |
| import hashlib |
| import json |
| import os |
| import statistics |
| import subprocess |
| import sys |
| from pathlib import Path |
|
|
| ROOT = Path(__file__).resolve().parent |
| BANDS = ['4k', '8k', '16k', '32k', '64k'] |
|
|
|
|
| def read_rows(path): |
| with path.open(encoding='utf-8') as stream: |
| return [json.loads(line) for line in stream if line.strip()] |
|
|
|
|
| def prepare_generations(rows): |
| """Accept case_id/mt; preserve supplied metadata, mark absent values unknown.""" |
| result = [] |
| for row in rows: |
| item = dict(row) |
| item.setdefault('status', 'ok') |
| if item['status'] == 'ok' and isinstance(item.get('mt'), str): |
| for key in ['finish_reason', 'input_tokens', 'output_tokens', 'cap_hit', |
| 'model_revision', 'generation_config_sha256']: |
| item.setdefault(key, None) |
| item.setdefault('output_sha256', hashlib.sha256(item['mt'].encode()).hexdigest()) |
| result.append(item) |
| return result |
|
|
|
|
| def five_band_macro(summary): |
| groups = summary['groups'].get('length_band', {}) |
| means = {band: groups.get(band, {}).get('segale_comet') for band in BANDS} |
| return {'value': statistics.fmean(means.values()) |
| if all(x is not None for x in means.values()) else None, |
| 'scale': [0, 1], 'band_means': means, |
| 'coverage': {band: {'scored': groups.get(band, {}).get('scored_cases', 0), |
| 'total': groups.get(band, {}).get('cases', 0)} for band in BANDS}} |
|
|
|
|
| def main(): |
| parser = argparse.ArgumentParser(description=__doc__) |
| parser.add_argument('--example', action='store_true') |
| parser.add_argument('--data-root', type=Path) |
| parser.add_argument('--generations', type=Path) |
| parser.add_argument('--output', type=Path, required=True, help='New output directory') |
| parser.add_argument('--device', choices=['cpu', 'cuda'], default='cpu') |
| parser.add_argument('--batch-size', type=int, default=8) |
| parser.add_argument('--case-workers', type=int, default=1) |
| args = parser.parse_args() |
| if args.example: |
| if args.data_root or args.generations: |
| parser.error('--example supplies both data and predictions') |
| args.data_root = ROOT / 'example' |
| args.generations = ROOT / 'example/generations.jsonl' |
| if not args.data_root or not args.generations: |
| parser.error('provide --example, or --data-root and --generations') |
| if args.batch_size < 1 or args.case_workers < 1: |
| parser.error('batch size and case workers must be positive') |
| if args.output.exists(): |
| parser.error('--output must be a new directory') |
| from huggingface_hub import hf_hub_download |
| models = json.loads((ROOT / 'models.json').read_text()) |
| resolved = {} |
| for name, spec in models.items(): |
| filename = 'checkpoints/model.ckpt' if name == 'comet' else 'config.json' |
| resolved[name] = Path(hf_hub_download(spec['repo_id'], filename, |
| revision=spec['revision'], local_files_only=True)) |
| args.output.mkdir(parents=True) |
| generations = args.output / 'generations.jsonl' |
| generations.write_text(''.join(json.dumps(r, ensure_ascii=False) + '\n' |
| for r in prepare_generations(read_rows(args.generations))), encoding='utf-8') |
| cases = args.data_root / 'cases.jsonl' |
| units = args.data_root / 'segale/alignment-units.jsonl' |
| adapter = args.output / 'adapter' |
| comet = args.output / 'comet' |
| chrf = args.output / 'chrf2' |
| env = os.environ.copy() |
| env['PYTHONPATH'] = os.pathsep.join([str(ROOT / 'vendor/segale'), str(ROOT / 'scripts')]) |
| env['HF_HUB_OFFLINE'] = '1' |
| env['TRANSFORMERS_OFFLINE'] = '1' |
| env['TOKENIZERS_PARALLELISM'] = 'false' |
| env.setdefault('OMP_NUM_THREADS', '4') |
| identity = ['--suite-id', 'nativelongbenchmark', '--system-key', 'prediction'] |
|
|
| def run(script, *values): |
| subprocess.run([sys.executable, str(ROOT / 'scripts' / script), |
| *map(str, values)], env=env, check=True) |
|
|
| run('score_document_chrf.py', '--data-manifest', args.data_root / 'manifest.json', |
| '--cases', cases, '--alignment-units', units, '--generations', generations, |
| '--output-dir', chrf, '--group-by', 'length_band', *identity) |
| run('prepare_document_segale.py', '--cases', cases, '--alignment-units', units, |
| '--generations', generations, '--output-dir', adapter, '--allow-incomplete', *identity) |
| selected = json.loads((adapter / 'manifest.json').read_text())['cases'] |
| if selected: |
| run('run_document_segale_alignment.py', '--system-file', adapter / 'system.jsonl', |
| '--ref-file', adapter / 'reference.jsonl', '--task-lang', 'en', |
| '--proc-device', args.device, '--embedding-model', resolved['bge'].parent, |
| '--max-size', '8', '--scratch-dir', args.output / 'scratch', |
| '--case-workers', args.case_workers, '--search-mode', 'online-stop') |
| run('evaluate_comet.py', '--input-file', adapter / 'system/aligned_spacy_system.jsonl', |
| '--manifest', adapter / 'manifest.json', '--output-dir', comet, |
| '--model-checkpoint', resolved['comet'], '--encoder-model', resolved['encoder'].parent, |
| '--batch-size', args.batch_size, '--gpus', int(args.device == 'cuda'), |
| '--spacy-model', 'en_core_web_sm', |
| '--target-sentences', adapter / 'system/target_sentences.jsonl') |
| else: |
| comet.mkdir() |
| (comet / 'summary.json').write_text(json.dumps({'cases': []}) + '\n') |
| run('summarize_document_segale.py', '--cases', cases, '--generations', generations, |
| '--comet-summary', comet / 'summary.json', '--output', args.output / 'summary.json', |
| '--group-by', 'work_id', '--group-by', 'length_band', '--chrf-dir', chrf, *identity) |
| summary = json.loads((args.output / 'summary.json').read_text()) |
| summary['five_band_macro'] = five_band_macro(summary) |
| summary['evaluation_profile'] = {'max_size': 8, 'segmenter': 'en_core_web_sm-3.8.0', |
| 'embedding_max_tokens': 512, 'search_mode': 'online-stop', 'device': args.device, |
| 'batch_size': args.batch_size, 'case_workers': args.case_workers, |
| 'models': {k: {'repo_id': v['repo_id'], 'revision': v['revision']} for k, v in models.items()}} |
| (args.output / 'summary.json').write_text(json.dumps(summary, ensure_ascii=False, indent=2) + '\n') |
| print(json.dumps({'case_count': summary['case_count'], 'scored': summary['scored_case_count'], |
| 'segale_comet': summary['scored']['segale_comet'], |
| 'five_band_macro': summary['five_band_macro']['value'], |
| 'summary': str(args.output / 'summary.json')}, indent=2)) |
|
|
|
|
| if __name__ == '__main__': |
| main() |
|
|
