"""Compare three frozen pretraining checkpoints on reviewed validation excerpts.""" import argparse import collections import gc import hashlib import json import math from pathlib import Path import sqlite3 import time import torch import torch.nn.functional as F from transformers import AutoConfig, AutoModelForCausalLM, GenerationConfig, set_seed from classical_tokenizer import ClassicalTokenizer ROOT = Path(__file__).resolve().parents[1] RUN = ROOT / 'runs/classical-smollm2-trunk' PACK = ROOT / 'unified/evaluation-tasks-v1/validation-001' CHECKPOINTS = {'anchor-24415': 'pre-decay-anchor-00024415', 'best-27500': 'checkpoint-00027500', 'final-30518': 'checkpoint-00030518'} # Decisions frozen before observing model outputs. Bounds are exact source substrings. SELECTION = [ (0, '料理手順', '染て網杓子', 'なかすがよし', None, '冒頭は前文からの続き。料理の操作を含むまとまった段落を与える。'), (3, '仏教説話', 'さつなれば。', '善女りうわうをくわんじやうし給ふ。', None, '冒頭は前文からの続き。祈雨の場面までを入力する。'), (12, '怪談・口語叙述', '正兵へ', 'のぞくまいそといつただけうすきみわろく', None, '翻刻の細かい改行を保存。覗くことを禁じられる文脈を与える。'), (14, '産業書の序文', '垝をしたゝめ', '蚕の患なからしめさせん事を欲するなり', '万延改元', '末尾の欠字記号を含む年記を対象範囲から除外。原文の曖昧な語は修正しない。'), (38, '平安随筆', '經など習ひて、', 'ふと覺ゆれ。', None, 'Wikisource見出しを除外し、本文の完結した一文を入力する。'), (62, '平安会話', 'かたはらなる人', '家の主と、定め申すべきことのはべるなり」といへば、', None, '発言と地の文を含む会話本文。38と同作品群のため独立8作品とは扱わない。'), (46, '軍記・候文', '但し当家の運命、', '強く御返答候処、', '【 NDLJP:', 'ページ標識の手前までを参照継続に使う。継続が短い例として明記する。'), (55, '漢文体の軍法', '一先陳之輩', '太不可然。', '以下闕脱', '欠脱表示の手前まで。原文の校訂注〔命歟〕は保存し、読解の曖昧さを明記する。'), ] def sha(path): h = hashlib.sha256() with open(path, 'rb') as f: for block in iter(lambda: f.read(1 << 20), b''): h.update(block) return h.hexdigest() def write_json(path, value): Path(path).write_text(json.dumps(value, ensure_ascii=False, indent=2, allow_nan=False) + '\n') def prepare(tokenizer): manifest = json.loads((PACK / 'manifest.json').read_text()) assert manifest['split'] == 'validation' assert sha(PACK / 'cases.jsonl') == manifest['files']['cases.jsonl'] assert sha(RUN / 'export/tokenizer.model') == manifest['tokenizer_sha256'] cases = [json.loads(x) for x in (PACK / 'cases.jsonl').read_text().splitlines()] split_rows = {r['id']: r for r in map(json.loads, (ROOT / 'unified/training-tokens-v1/build-006/splits.jsonl').read_text().splitlines())} db = sqlite3.connect('file:' + str(ROOT / 'unified/pretrain-v1/build-002/build.sqlite3') + '?mode=ro', uri=True) selected = [] try: for index, style, begin, end, stop, note in SELECTION: case = cases[index] assert split_rows[case['source_id']]['split'] == 'validation' assert split_rows[case['source_id']]['group'] == case['group'] text = db.execute('SELECT text FROM units WHERE id=?', (case['source_id'],)).fetchone()[0] assert hashlib.sha256(text.encode()).hexdigest() == case['text_sha256'] assert text[case['span_start']:case['span_end']] == case['raw_episode'] raw = case['raw_episode'] start = raw.index(begin) boundary = raw.index(end, start) + len(end) finish = raw.index(stop, boundary) if stop else len(raw) prompt = raw[start:boundary] reference = raw[boundary:finish] assert len(reference) >= 24 assert not any(x in prompt + reference for x in ('■', '□', '〓', '【 NDLJP:', '==')) ids = tokenizer.raw_encode(prompt + reference) assert tokenizer.decode(ids) == prompt + reference selected.append(dict(id=f'case-{len(selected)+1:02d}', fixture_index=index, fixture_id=case['id'], title=case['title'], source=case['source'], group=case['group'], era=case['era'], style=style, source_id=case['source_id'], source_text_sha256=case['text_sha256'], source_start=case['span_start'] + start, source_end=case['span_start'] + finish, prompt=prompt, reference=reference, review_note=note, review_status='Codex reviewed selected excerpt; not expert certification', prompt_tokens=len(tokenizer.raw_encode(prompt)) + 1)) finally: db.close() return selected def repetition(ids, n=4): grams = [tuple(ids[i:i+n]) for i in range(len(ids)-n+1)] return 1 - len(set(grams)) / len(grams) if grams else 0. @torch.inference_mode() def score(model, tokenizer, case): # Whole fixed raw excerpt, BOS included as input but never a prediction target. text = case['prompt'] + case['reference'] ids = torch.tensor([[1] + tokenizer.raw_encode(text)], device='cuda') with torch.autocast('cuda', dtype=torch.bfloat16): logits = model(input_ids=ids[:, :-1], use_cache=False).logits total = float(F.cross_entropy(logits.float().reshape(-1, logits.shape[-1]), ids[:, 1:].reshape(-1), reduction='sum')) return dict(raw_characters=len(text), predicted_tokens=ids.shape[1]-1, nll_sum=total, nll=total/(ids.shape[1]-1), bpc=total/(math.log(2)*len(text))) def summary(rows, scores): result = {} for checkpoint in CHECKPOINTS: s = [r for r in scores if r['checkpoint'] == checkpoint] by_source = {} for source in ('minhon', 'wikisource'): chosen = [r for r in s if r['source'] == source] by_source[source] = sum(r['nll_sum'] for r in chosen) / (math.log(2)*sum(r['raw_characters'] for r in chosen)) result[checkpoint] = dict(by_source_bpc=by_source, equal_source_bpc=sum(by_source.values())/len(by_source), modes={}) for mode in ('greedy', 'sampling'): rs = [r for r in rows if r['checkpoint'] == checkpoint and r['mode'] == mode] result[checkpoint]['modes'][mode] = dict(examples=len(rs), eos_stops=sum(r['stopped_with_eos'] for r in rs), mean_repeated_token_4gram_fraction=sum(r['repeated_token_4gram_fraction'] for r in rs)/len(rs), mean_generated_tokens=sum(len(r['generated_ids']) for r in rs)/len(rs)) return result def report(output, cases, rows, scores, stats): lines = ['# 事前学習3 checkpointの生成比較', '', 'validationから事前に選んだ8例・7作品群。学習やtest評価は行っていない。選択範囲はCodexが目視確認し、原文位置・SHA256・splitを照合した。既存64例全体の専門家点検を完了したものではない。', '', '本文続きを生成するbaseモデルとして、同一prompt・BOS・既存tokenizerを使用。各生成は最大128新規token。greedyとtemperature=0.8/top_p=0.95のsamplingを比較。samplingは例ごとに同一seedを3モデルへ設定するが、各モデルの確率分布は異なる。反復抑制の追加設定は入れていない。', '', '| checkpoint | 等出典平均BPC | greedy反復4gram比 | sampling反復4gram比 |', '|---|---:|---:|---:|'] for name, row in stats.items(): lines.append(f"| {name} | {row['equal_source_bpc']:.4f} | {row['modes']['greedy']['mean_repeated_token_4gram_fraction']:.3f} | {row['modes']['sampling']['mean_repeated_token_4gram_fraction']:.3f} |") lines += ['', 'BPCはpromptと参照継続を含む同一原文範囲全体のteacher-forced指標で、低いほどよい。反復比は生成tokenの4gramが既出である割合の平均。自然な繰り返しも数えるため品質判定そのものではない。8例の同一作品群を含む小規模比較で、有意差や汎化を保証しない。', 'EOS終了数はsummary.jsonに保存。baseモデルには本文を続ける役割があるため、EOSが出ないこと自体を失敗とはしない。原文参照の完全再現も成功条件ではなく、文体・話題・文章のつながりを併せて読む。', ''] for case in cases: lines += [f"## {case['id']} {case['title']} — {case['style']}", '', f"出典: {case['source']} / {case['era']} / {case['source_id']}。prompt {case['prompt_tokens']} token。", '', '点検メモ: ' + case['review_note'], '', '### 入力', '', '```text', case['prompt'], '```', '', '### 原文の参照継続', '', '```text', case['reference'], '```', ''] for mode in ('greedy', 'sampling'): lines += ['### ' + mode, ''] for name in CHECKPOINTS: row = next(r for r in rows if r['id'] == case['id'] and r['mode'] == mode and r['checkpoint'] == name) lines += [f"**{name}**（{len(row['generated_ids'])} token、EOS={row['stopped_with_eos']}、反復比={row['repeated_token_4gram_fraction']:.3f}）", '', '```text', row['generated'], '```', ''] (output / 'COMPARISON.md').write_text('\n'.join(lines)) def main(): parser = argparse.ArgumentParser(description=__doc__) parser.add_argument('--output', default='runs/pretrain-generation-comparison-001') args = parser.parse_args() output = Path(args.output) output.mkdir(parents=True, exist_ok=False) torch.set_num_threads(4) tokenizer = ClassicalTokenizer(RUN / 'export/tokenizer.model') cases = prepare(tokenizer) write_json(output / 'cases.json', cases) model_hashes = {name: sha(RUN / 'checkpoints' / directory / 'model.safetensors') for name, directory in CHECKPOINTS.items()} write_json(output / 'manifest.json', dict(split='validation', seed_rule='20261027 + case_index', max_new_tokens=128, temperature=0.8, top_p=0.95, checkpoints=CHECKPOINTS, checkpoint_sha256=model_hashes, tokenizer_sha256=sha(RUN / 'export/tokenizer.model'), fixture_sha256=sha(PACK / 'cases.jsonl'), script_sha256=sha(__file__), case_count=len(cases), independent_groups=len({r['group'] for r in cases}), case_file_sha256=sha(output / 'cases.json'), device='cuda', weights='float32', autocast='bfloat16')) scores, rows = [], [] start = time.monotonic() for name, directory in CHECKPOINTS.items(): model = AutoModelForCausalLM.from_pretrained(RUN / 'checkpoints' / directory, config=AutoConfig.from_pretrained(RUN / 'export', local_files_only=True), generation_config=GenerationConfig.from_pretrained(RUN / 'export', local_files_only=True), local_files_only=True, attn_implementation='sdpa', dtype=torch.float32).cuda().eval() for index, case in enumerate(cases): scores.append(dict(id=case['id'], checkpoint=name, source=case['source'], **score(model, tokenizer, case))) ids = torch.tensor([[1] + tokenizer.raw_encode(case['prompt'])], device='cuda') for mode in ('greedy', 'sampling'): set_seed(20261027 + index) config = GenerationConfig(max_new_tokens=128, do_sample=mode == 'sampling', eos_token_id=2, pad_token_id=3, bos_token_id=1, use_cache=True) if mode == 'sampling': config.temperature, config.top_p = 0.8, 0.95 with torch.inference_mode(), torch.autocast('cuda', dtype=torch.bfloat16): generated = model.generate(input_ids=ids, attention_mask=torch.ones_like(ids), generation_config=config) new = generated[0, ids.shape[1]:].tolist() row = dict(id=case['id'], checkpoint=name, source=case['source'], mode=mode, seed=20261027 + index, generated_ids=new, generated=tokenizer.decode(new, skip_special_tokens=True), stopped_with_eos=bool(new and new[-1] == 2), repeated_token_4gram_fraction=repetition([i for i in new if i not in (1, 2, 3)])) rows.append(row) with (output / 'generations.jsonl').open('a') as f: f.write(json.dumps(row, ensure_ascii=False, allow_nan=False) + '\n') print(json.dumps(dict(checkpoint=name, case=case['id'], completed_generations=len(rows), elapsed_seconds=time.monotonic()-start)), flush=True) del model gc.collect() torch.cuda.empty_cache() stats = summary(rows, scores) write_json(output / 'scores.json', scores) write_json(output / 'summary.json', stats) report(output, cases, rows, scores, stats) assert len(rows) == 48 and len(scores) == 24 unchanged = {name: sha(RUN / 'checkpoints' / directory / 'model.safetensors') for name, directory in CHECKPOINTS.items()} assert unchanged == model_hashes write_json(output / 'verification.json', dict(complete=True, generations=len(rows), scores=len(scores), checkpoints_unchanged=True, cases_verified_against_validation_source=True, artifact_sha256={p.name: sha(p) for p in output.iterdir() if p.is_file()}, elapsed_seconds=time.monotonic()-start)) if __name__ == '__main__': main()