Blame for code/scripts/audit_tokenization_efficiency.py

"""Read-only efficiency audit of frozen vocabulary and source/work-balanced text samples."""
import argparse
import collections
import hashlib
import json
import math
from pathlib import Path
import sqlite3
import time
import unicodedata

from sentencepiece_unigram import Tokenizer

ROOT = Path(__file__).resolve().parents[1]
DATA = ROOT / 'unified/training-tokens-v1/build-006'
MODEL = ROOT / 'runs/classical-smollm2-trunk/export/tokenizer.model'


def sha(path):
    h = hashlib.sha256()
    with open(path, 'rb') as f:
        for b in iter(lambda: f.read(1 << 20), b''):
            h.update(b)
    return h.hexdigest()


def write(path, value):
    Path(path).write_text(json.dumps(value, ensure_ascii=False, indent=2, allow_nan=False) + '\n')


def rank(value):
    return hashlib.sha256(('efficiency-20261027:' + value).encode()).hexdigest()


def category(text):
    if text.isspace():
        return 'whitespace'
    if all(unicodedata.category(c)[0] in 'PS' for c in text):
        return 'punctuation_symbols'
    if all(('\u3400' <= c <= '\u9fff') or ('\U00020000' <= c <= '\U000323af') or ('\uf900' <= c <= '\ufaff') for c in text):
        return 'kanji_only'
    if all('\u3040' <= c <= '\u30ff' for c in text):
        return 'kana_only'
    if any('\u3040' <= c <= '\u30ff' for c in text):
        return 'contains_kana_mixed'
    return 'other'


class MinimumTokens:
    """Exact shortest reversible token sequence with current regular/byte pieces.

    This is a compression bound, not the probability-optimal Unigram objective.
    Whitespace uses the model's existing identity normalizer and U+2581 rule.
    """
    def __init__(self, tokenizer):
        self.tokenizer = tokenizer
        self.trie = {}
        sp = tokenizer.sp
        self.byte_ids = [sp.piece_to_id(f'<0x{i:02X}>') for i in range(256)]
        for i in range(sp.vocab_size()):
            if sp.is_byte(i) or sp.is_control(i) or sp.is_unknown(i) or sp.is_unused(i):
                continue
            surface = sp.id_to_piece(i).replace('▁', ' ')
            if not surface:
                continue
            node = self.trie
            for c in surface:
                node = node.setdefault(c, {})
            node[None] = i

    def encode(self, text):
        n = len(text)
        costs = [10**9] * (n + 1)
        previous = [None] * (n + 1)
        costs[0] = 0
        for start in range(n):
            # Every Unicode scalar has a reversible UTF-8 byte fallback.
            fallback = [self.byte_ids[b] for b in text[start].encode()]
            if costs[start] + len(fallback) < costs[start + 1]:
                costs[start + 1] = costs[start] + len(fallback)
                previous[start + 1] = (start, fallback)
            node = self.trie
            end = start
            while end < n and text[end] in node:
                node = node[text[end]]
                end += 1
                if None in node and costs[start] + 1 < costs[end]:
                    costs[end] = costs[start] + 1
                    previous[end] = (start, [node[None]])
        parts, end = [], n
        while end:
            begin, ids = previous[end]
            parts.append(ids)
            end = begin
        ids = [i for part in reversed(parts) for i in part]
        if self.tokenizer.decode(ids) != text:
            raise ValueError('Minimum-token path does not round-trip')
        return ids


def select():
    groups = collections.defaultdict(list)
    with (DATA / 'splits.jsonl').open() as f:
        for line in f:
            row = json.loads(line)
            if row['split'] in ('train', 'validation'):
                groups[row['source'], row['split'], row['group']].append(row)
    by_partition = collections.defaultdict(list)
    for (source, split, group), rows in groups.items():
        by_partition[source, split].append((group, min(rows, key=lambda r: rank(r['id']))))
    selected = []
    for partition in sorted(by_partition):
        works = sorted(by_partition[partition], key=lambda p: rank(p[0]))[:128]
        selected.extend(row for _, row in works)
    return selected


def summarize(rows):
    counts = collections.Counter()
    for row in rows:
        for key in ('characters', 'nonwhitespace_characters', 'tokens', 'regular_tokens',
                    'one_character_regular_tokens', 'one_character_content_tokens',
                    'content_regular_tokens', 'byte_tokens', 'newline_byte_tokens', 'unknown_tokens',
                    'minimum_tokens', 'no_lf_tokens', 'newline_characters'):
            counts[key] += row[key]
    chars, tokens = counts['characters'], counts['tokens']
    cpt = sorted(r['characters']/r['tokens'] for r in rows)
    return dict(samples=len(rows), works=len({r['group'] for r in rows}), **dict(counts),
        characters_per_token=chars/tokens, tokens_per_100_characters=100*tokens/chars,
        median_document_characters_per_token=cpt[len(cpt)//2],
        p10_document_characters_per_token=cpt[int((len(cpt)-1)*.1)],
        p90_document_characters_per_token=cpt[int((len(cpt)-1)*.9)],
        one_character_regular_fraction=counts['one_character_regular_tokens']/max(1, counts['regular_tokens']),
        one_character_content_fraction=counts['one_character_content_tokens']/max(1, counts['content_regular_tokens']),
        byte_token_fraction=counts['byte_tokens']/tokens,
        non_lf_byte_fraction=(counts['byte_tokens']-counts['newline_byte_tokens'])/tokens,
        minimum_token_reduction_fraction=1-counts['minimum_tokens']/tokens,
        no_lf_token_reduction_fraction=1-counts['no_lf_tokens']/tokens)


def main():
    parser = argparse.ArgumentParser(description=__doc__)
    parser.add_argument('--output', default='runs/tokenization-efficiency-audit-002')
    args = parser.parse_args()
    output = ROOT / args.output
    output.mkdir(exist_ok=False)
    started = time.monotonic()
    tok = Tokenizer(MODEL)
    sp = tok.sp
    manifest = json.loads((DATA / 'manifest.json').read_text())
    assert sha(MODEL) == manifest['tokenizer_sha256']
    optimal = MinimumTokens(tok)
    selected = select()
    rows = []
    usage = collections.defaultdict(collections.Counter)
    examples = []
    db = sqlite3.connect('file:' + str(ROOT / 'unified/pretrain-v1/build-002/build.sqlite3') + '?mode=ro', uri=True)
    newline_id = sp.piece_to_id('<0x0A>')
    with (output / 'samples.jsonl').open('w') as f:
        for index, r in enumerate(selected):
            document = db.execute('SELECT text FROM units WHERE id=?', (r['id'],)).fetchone()[0]
            assert hashlib.sha256(document.encode()).hexdigest() == r['text_sha256']
            start = max(0, (len(document)-2048)//2)
            text = document[start:start+2048]
            ids = tok.encode(text)
            assert ids and tok.decode(ids) == text
            shortest = optimal.encode(text)
            assert len(shortest) <= len(ids)
            no_lf = tok.encode(text.replace('\n', ''))
            regular = [sp.id_to_piece(i).replace('▁', ' ') for i in ids if not sp.is_byte(i) and not sp.is_control(i) and not sp.is_unknown(i)]
            content = [p for p in regular if category(p) not in ('whitespace', 'punctuation_symbols')]
            row = dict(source=r['source'], split=r['split'], id=r['id'], group=r['group'],
                title=r['title'], era=r['era'], text_sha256=r['text_sha256'], start=start, end=start+len(text),
                characters=len(text), nonwhitespace_characters=sum(not c.isspace() for c in text), tokens=len(ids),
                regular_tokens=len(regular), one_character_regular_tokens=sum(len(p)==1 for p in regular),
                content_regular_tokens=len(content), one_character_content_tokens=sum(len(p)==1 for p in content),
                byte_tokens=sum(sp.is_byte(i) for i in ids), newline_byte_tokens=ids.count(newline_id),
                unknown_tokens=ids.count(0), minimum_tokens=len(shortest), no_lf_tokens=len(no_lf),
                newline_characters=text.count('\n'), roundtrip_exact=True)
            rows.append(row)
            f.write(json.dumps(row, ensure_ascii=False) + '\n')
            usage[r['source']].update(ids)
            if r['source'] == 'wikisource':
                examples.append(dict(id=r['id'], split=r['split'], title=r['title'],
                    original_tokens=len(ids), minimum_tokens=len(shortest),
                    gain=1-len(shortest)/len(ids), text=text[:256],
                    first_128_characters=text[:128],
                    current_pieces=[sp.id_to_piece(i) for i in tok.encode(text[:128])],
                    minimum_pieces=[sp.id_to_piece(i) for i in optimal.encode(text[:128])]))
            if (index+1) % 100 == 0:
                print(json.dumps(dict(samples=index+1, total=len(selected), seconds=time.monotonic()-started)), flush=True)
    db.close()
    report = {'by_source': {}, 'by_source_split': {}, 'wikisource_by_era': {}}
    for source in sorted({r['source'] for r in rows}):
        report['by_source'][source] = summarize([r for r in rows if r['source']==source])
        for split in ('train', 'validation'):
            rs = [r for r in rows if r['source']==source and r['split']==split]
            if rs:
                report['by_source_split'][source+'/'+split] = summarize(rs)
    for era in sorted({r['era'] for r in rows if r['source']=='wikisource'}):
        report['wikisource_by_era'][era] = summarize([r for r in rows if r['source']=='wikisource' and r['era']==era])
    # Keep all fixed, previously reviewed styles as a separate diagnostic, not pooled into work samples.
    styles = json.loads((ROOT / 'runs/pretrain-generation-comparison-001/cases.json').read_text())
    report['reviewed_styles'] = []
    for case in styles:
        text = case['prompt'] + case['reference']
        ids = tok.encode(text)
        shortest = optimal.encode(text)
        report['reviewed_styles'].append(dict(id=case['id'], source=case['source'], title=case['title'],
            style=case['style'], characters=len(text), tokens=len(ids), minimum_tokens=len(shortest),
            characters_per_token=len(text)/len(ids), minimum_token_reduction_fraction=1-len(shortest)/len(ids)))
    vocab = collections.Counter()
    for i in range(sp.vocab_size()):
        if sp.is_byte(i):
            vocab['byte'] += 1
        elif sp.is_control(i) or sp.is_unknown(i):
            vocab['special'] += 1
        else:
            surface = sp.id_to_piece(i).replace('▁', ' ')
            vocab[category(surface)] += 1
            vocab['single_character' if len(surface)==1 else 'multiple_characters'] += 1
    sampling = json.loads((ROOT / 'unified/tokenizer-v1/build-001/sampling.json').read_text())
    total = sum(r['sampled_characters'] for r in sampling['by_source'].values())
    report['tokenizer_training_sample_character_shares'] = {s:r['sampled_characters']/total for s,r in sampling['by_source'].items()}
    report['vocabulary_categories_overlapping_length_counts'] = dict(vocab)
    report['top_one_character_content_pieces'] = {}
    for source, counts in usage.items():
        report['top_one_character_content_pieces'][source] = [dict(piece=sp.id_to_piece(i), id=i, count=n)
            for i,n in counts.most_common() if not sp.is_byte(i) and len(sp.id_to_piece(i))==1
            and category(sp.id_to_piece(i)) not in ('whitespace', 'punctuation_symbols')][:30]
    write(output / 'report.json', report)
    write(output / 'wikisource-examples.json', sorted(examples, key=lambda e:e['gain'], reverse=True))
    write(output / 'manifest.json', dict(complete=True, source_splits=['train','validation'], test_read=False,
        selection='128 groups per source/split by fixed SHA256 rank, one document per group, middle up to 2048 Unicode characters',
        tokenizer_sha256=sha(MODEL), script_sha256=sha(__file__), splits_sha256=sha(DATA/'splits.jsonl'),
        samples=len(rows), independent_global_groups=len({r['group'] for r in rows}),
        roundtrip_matches=len(rows), unknown_tokens=sum(r['unknown_tokens'] for r in rows),
        minimum_paths_verified=len(rows), elapsed_seconds=time.monotonic()-started,
        caveat='Tokenizer predates LLM splitting; samples are not tokenizer-unseen. Compression bound does not optimize language-model quality.'))
    lines = ['# トークナイズ分割効率の診断', '',
        f'{len(rows)}出典内作品サンプル（独立global groupは{len({r["group"] for r in rows})}）から各1レコードの中央最大2,048文字を採用。train/validationのみ。tokenizerは作品分割より前に作成済みなので、tokenizer未知作品への評価ではない。', '',
        '| 出典 | 作品数 | 文字/token ↑ | 100文字のtoken数 ↓ | 内容token中の1文字比 | byte比 | byte比（LF以外） | 同一語彙での最短化削減率 |',
        '|---|---:|---:|---:|---:|---:|---:|---:|']
    for s,r in report['by_source'].items():
        lines.append(f"| {s} | {r['samples']} | {r['characters_per_token']:.3f} | {r['tokens_per_100_characters']:.1f} | {r['one_character_content_fraction']:.1%} | {r['byte_token_fraction']:.2%} | {r['non_lf_byte_fraction']:.2%} | {r['minimum_token_reduction_fraction']:.3%} |")
    lines += ['', '文字/tokenはUnicode文字数（空白・改行を含む）÷本文token数。BOS/EOS/PADを除外。内容tokenは通常語彙のうち空白・句読点・記号だけのpieceを除いたもの。語の正しさを測る指標ではない。', '',
        '最短化は現行32,000語彙の通常pieceとUTF-8 byte piecesだけで原文を完全復元する最少token列を動的計画法で計算した比較上限。全サンプルで復元一致を確認した。Unigramの確率を最大化する通常encodeとは目的が違い、最短分割を既存LLMへそのまま入力する提案ではない。', '',
        '## train/validation別', '', '| 区分 | 作品数 | 文字/token | 同一語彙最短化削減率 |', '|---|---:|---:|---:|']
    for k,r in report['by_source_split'].items():
        lines.append(f"| {k} | {r['samples']} | {r['characters_per_token']:.3f} | {r['minimum_token_reduction_fraction']:.3%} |")
    lines += ['', '## 固定文体例', '', '| 作品・文体 | 文字数 | token数 | 文字/token | 同一語彙最短化削減率 |', '|---|---:|---:|---:|---:|']
    for r in report['reviewed_styles']:
        lines.append(f"| {r['title']}・{r['style']} | {r['characters']} | {r['tokens']} | {r['characters_per_token']:.3f} | {r['minimum_token_reduction_fraction']:.3%} |")
    lines += ['', '## 改行の診断', '', '| 出典 | LF文字数 | LF削除によるtoken数削減率 |', '|---|---:|---:|']
    for s,r in report['by_source'].items():
        lines.append(f"| {s} | {r['newline_characters']} | {r['no_lf_token_reduction_fraction']:.1%} |")
    lines += ['', 'LF削除はtoken負荷の感度診断だけであり、原文を変更する操作。語の結合による変化も含む。実データには適用せず、恣意的な改行の一律削除は提案しない。', '',
        '## tokenizer学習時の出典配分（文字数）', '']
    for s,v in report['tokenizer_training_sample_character_shares'].items():
        lines.append(f'- {s}: {v:.2%}')
    lines += ['', 'NDLは和文・漢文等を含むため、出典だけから言語比率とは見なさない。', '',
        '## 語彙と使用例', '', '```json', json.dumps(dict(vocab), ensure_ascii=False, indent=2), '```', '',
        '語彙の長さ別カウントは文字種別と重複する。文字種判定は簡易Unicode規則で、形態素や言語の厳密判定ではない。', '',
        'report.jsonに出典別頻出1文字piece、Wikisourceの時代別数値を保存。wikisource-examples.jsonには同一原文での通常分割と最短分割を保存。サンプルの作品・位置・原文hash・数値はsamples.jsonl。', '',
        '## 解釈の限界', '',
        '文字/tokenだけで他言語・異なるtokenizerの優劣やモデル品質は判断しない。同語彙最短化は語彙固定での圧縮余地を示すが、より良い語彙を再学習した場合の改善幅は未測定。既存モデル・tokenizer・学習用binは変更していない。', '']
    (output/'REPORT.md').write_text('\n'.join(lines))
    print(json.dumps(report['by_source'], ensure_ascii=False, indent=2))


if __name__ == '__main__':
    main()