processing/phase6.yaml

extends: configs/base.yaml
phase: phase6
# Phase 6 reopens the corpus. Phase 5 concluded that this closed corpus was near
# an entropy floor because two single-lever moves (data 2.5x, model 2.1x) both
# bought ~3.5% BPB. reports/phase6_rejection_review.json shows that conclusion
# rested on a corpus that was never exhausted: 141,521 of the frozen pool's
# 241,430 candidates (58.6%) were dropped as short_document, and the filter
# responsible was measuring domain novelty rather than OCR corruption -- its
# reference 3-gram set came from the Aozora literary corpus alone (2.88M
# 3-grams). Re-scoring 300 re-downloaded quarantined documents against a
# reference that also includes accepted NDL text recovers 63% of them, worth an
# estimated +1.08B prose tokens (train 2.742B -> 3.818B).
#
# 6.2 (this config's acquisition stage) only re-downloads. Cleaning waits for
# the 6.1 filter redesign, so raw ZIPs are NOT pruned in between -- pruning
# before the new filter exists would force a second five-day re-download.
data:
  ndl_metadata_dir: data/raw/ndl/metadata
  ndl_metadata_glob: data/raw/ndl/metadata/books_*.zip
  ndl_metadata_urls:
    - https://dl.ndl.go.jp/static/files/dataset/dataset_202602_t_internet_01.zip
    - https://dl.ndl.go.jp/static/files/dataset/dataset_202602_t_internet_02.zip
    - https://dl.ndl.go.jp/static/files/dataset/dataset_202602_t_internet_03.zip
    - https://dl.ndl.go.jp/static/files/dataset/dataset_202602_t_internet_04.zip
    - https://dl.ndl.go.jp/static/files/dataset/dataset_202602_t_internet_05.zip
  # The frozen pool stays the source of truth for order and rights evidence.
  ndl_selection_pool_manifest: data/processed/ndl/selection.jsonl
  # ndl-download walks ndl_selection_manifest, so phase 6 points it at the
  # quarantined-only subset (in pool order) built by
  # scripts/build_phase6_selection.py. Existing ZIPs are skipped by checksum, so
  # the run stays idempotent and resumable. Do NOT run ndl-select against this
  # config: the pool is frozen and ndl-select would overwrite this subset.
  ndl_selection_manifest: data/processed/phase6/selection_quarantined.jsonl
  ndl_download_manifest: data/processed/phase6/ndl_downloads.jsonl
  ndl_raw_dir: data/raw/ndl/books
  ndl_clean_dir: data/clean/ndl
  ndl_clean_manifest: data/processed/phase6/ndl_clean.jsonl
  # Phase-5's manifest is already a superset carrying phases 1-4 forward.
  ndl_clean_reuse_manifest: data/processed/phase5/ndl_clean.jsonl
  ndl_clean_report: phase6_ndl_cleaning.json
  ndl_quality_sample_report: phase6_ndl_quality_sample.jsonl
  ndl_quality_exclude_manifest: data/processed/phase5/ndl_clean.jsonl
  ndl_target_documents: 241430
  ndl_download_delay_seconds: 0.2
  ndl_download_retries: 3
  ndl_max_download_seconds: 600
  # 6.3 froze phase-6 splits, so this now points at them (phase-4's manifest stays
  # on disk untouched). The fixed 24-work phase-1 test set carries over unchanged.
  split_manifest: data/processed/phase6/splits.jsonl
  ndl_max_abnormal_ngram_rate: 0.38
  ndl_max_latin_rate: 0.005
  # 6.1 filter redesign. The abnormal-ngram reference was built from the Aozora
  # literary corpus alone (2,879,238 3-grams), which made that filter a
  # domain-novelty detector rather than an OCR-corruption one: re-scoring 300
  # re-downloaded quarantined documents against a reference that also includes
  # accepted NDL text recovers 63% of them (reports/phase6_rejection_review.json).
  # Order is fixed and each manifest is capped, so the reference itself is
  # reproducible and can be frozen alongside the thresholds.
  ndl_reference_manifests:
    - data/processed/aozora/clean.jsonl
    - data/processed/phase5/ndl_clean.jsonl
  ndl_reference_max_chars: 200000000
  # Widening the reference lets statistics tables and directories through, since
  # abnormal_ngram_rate had been rejecting them as a side effect. Calibrated in
  # reports/phase6_filter_calibration.json: a 0.18 numeral-run ratio fires on
  # 0.071% of already-accepted characters (inside 6.1's 0.5% shrink gate) and
  # catches 100% of the pages the audit triaged as tables.
  #
  # No kana-run rule ships. The draft "garbled" detector was measured against the
  # same two sets and no run length separates ruby-contaminated OCR from ordinary
  # 文語体 (run>=8 at ratio>0.05 flags 54% of accepted characters; run>=20 at
  # ratio>0.20 still flags 8% while recall falls to 41%). 文語体 simply contains
  # long kana runs. mojibake_page continues to handle genuine corruption.
  ndl_max_digit_run_rate: 0.18
  # 6.3 の phase6-merge が書き出す、ページ規則適用後のcleanテキスト。
  # data/clean/ndl を直接書き換えない: 記録済み clean_sha256 と公開済み v1/v2 の
  # 基礎が変わるため（reports/phase6_cleaning_plan.json で却下した案）。
  ndl_filtered_dir: data/clean/ndl_phase6
  # 仮名3-gram参照（青空文庫のみ、凍結）。kana_soup の判定に使う。
  # 生成: uv run python scripts/calibrate_phase6_kana_soup.py
  ndl_kana_reference: artifacts/phase6/kana_reference.txt
  # アプリの接地検索が引く母集団。ndl_clean_manifest（ページ規則適用前の
  # data/clean/ndl を指す）ではなく、モデルが実際に学習した ndl_phase6 側を
  # 向ける。書名索引に必要な項目（ndl_pid/title/ndc/publication_year/
  # clean_path/clean_chars）はsplit manifestが全て持っている。
  ndl_grounding_manifest: data/processed/phase6/splits.jsonl
tokenizer:
  # Retrained in 6.3 over the expanded corpus. BPB is per byte, so changing the
  # tokenizer does not break comparability with phases 1-5.
  vocab_size: 16384
  directory: artifacts/phase6/tokenizer
  max_train_chars: 200000000
tokenized:
  directory: data/processed/phase6/tokenized
model:
  # 6.4 co-scaling. フェーズ5は730M（36層/20head/1280）でデータ据え置き、
  # フェーズ4は345Mでデータ2.5倍。単独レバーの改善率がほぼ一致した
  # （+3.46% / +3.53%）ことがフェーズ6の出発点なので、ここは初めて両方を同時に
  # 動かす: データ 2.742B → 実測3.8B級、モデル 730M → 約1.05B。
  # head_dim = 1536/24 = 64 でフェーズ3〜5と同じ比を保つ。
  # 実数は `tcja --config configs/phase6.yaml parameter-count --vocab-size 16384`
  # で確認し reports/runs/parameter-count.json へ残す。
  n_layer: 36
  n_head: 24
  n_embd: 1536
  block_size: 1024
  dropout: 0.0
training:
  device: cuda
  dtype: bfloat16
  # ユーザー判断(2026-07-30)で batch16 x accum16。tokens_per_step は 262,144 で
  # 変わらないので max_iters もそのまま。
  # 当初は batch8 x accum32 を第一候補にしていたが、accumulation回数が半分になる
  # 分こちらが速いはずで、本学習163時間に対しては数%でも十数時間効く。
  # フェーズ5は同じ形（batch16 x accum16）を730Mで走らせて peak 43.8GB・余裕48.4%
  # だった。パラメータは1.43倍だが、支配的なoptimizer stateでも +9GB程度の見込み。
  # **6.5のflightで実測する。VRAM余裕が10%未満なら batch8 x accum32 へ戻す。**
  batch_size: 16
  gradient_accumulation_steps: 16
  # FINALIZED (2026-07-30、reports/phase6_tokenization.json): train実測
  # 3,953,739,596 tokens × 3ep = 11,861,218,788 処理トークン ÷ 262,144 tok/step
  # = 45,254.9 → 45,300 iters（3.003ep、処理 11,875,123,200）。
  # tokens/param 11.35 は フェーズ5の11.3とほぼ同じで、データとモデルを同率で
  # 伸ばした結果になっている。
  max_iters: 45300
  learning_rate: 0.00025
  min_lr: 0.000025
  warmup_iters: 300
  lr_decay_iters: 45300
  eval_interval: 500
  eval_iters: 50
  checkpoint_interval: 500
  grad_clip: 1.0
  out_dir: artifacts/phase6
budget:
  # フェーズ5実測 9.065 s/step（730M・A100 80GB・262,144 tok/step）= 104.1M tok/h。
  # FLOPsはパラメータ数にほぼ比例するので 1.05B では約72M tok/h。
  # 3ep（約11.4B処理）で約158h・$220（$1.39/h）。上限は余裕を見て置く。
  minimum_train_tokens: 3500000000
  maximum_hours: 190
  maximum_usd: 270
  gpu_hourly_usd: 1.39