Datasets:
| extends: configs/base.yaml | |
| phase: phase4 | |
| # Phase 4 keeps the 345M model frozen from phase 3 and only expands unique data. | |
| # Rationale (2026-07-18): phase-3 was data-bound, not capacity-bound | |
| # (1.103B unique train tokens = ~3.2 tokens/param vs Chinchilla-optimal ~7B for | |
| # 345M; small train/val gap; val loss still falling at LR floor). The next lever | |
| # is more unique tokens, not more parameters. Target case B below aims at ~3.0B | |
| # train tokens (~2.7x). The frozen NDL selection pool (241,430 candidates) tops | |
| # out near ~3.8B tokens under the current filters; beyond that needs new sources. | |
| 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 | |
| # Same frozen 241,430-candidate pool and selection order as phase 3. Expanding | |
| # the download target just consumes deeper into this manifest; the head 50,000 | |
| # zips are already on disk and skipped by checksum (never re-fetched). | |
| ndl_selection_manifest: data/processed/ndl/selection.jsonl | |
| ndl_download_manifest: data/processed/phase4/ndl_downloads.jsonl | |
| ndl_raw_dir: data/raw/ndl/books | |
| ndl_clean_dir: data/clean/ndl | |
| ndl_clean_manifest: data/processed/phase4/ndl_clean.jsonl | |
| # Reuse phase-3's 27,725 already-cleaned documents so ndl-clean only processes | |
| # the newly downloaded candidates. | |
| ndl_clean_reuse_manifest: data/processed/phase3/ndl_clean.jsonl | |
| ndl_clean_report: phase4_ndl_cleaning.json | |
| ndl_quality_sample_report: phase4_ndl_quality_sample.jsonl | |
| # Exclude phase-3 documents from the fresh 200-page human review sample. | |
| ndl_quality_exclude_manifest: data/processed/phase3/ndl_clean.jsonl | |
| # Case B (recommended). Head 50,000 already downloaded; this pulls ~136,000 | |
| # more (~14k tokens/candidate in the deep pool) toward ~3.0B train tokens. | |
| # Case A (fast start): 114000 -> ~2.0B. Case C (exhaust pool): 241430 -> ~3.8B. | |
| # The actual target is passed on the CLI: ndl-download --max-documents 186000. | |
| ndl_target_documents: 186000 | |
| ndl_download_delay_seconds: 0.2 | |
| ndl_download_retries: 3 | |
| ndl_max_download_seconds: 600 | |
| split_manifest: data/processed/phase4/splits.jsonl | |
| ndl_max_abnormal_ngram_rate: 0.38 | |
| ndl_max_latin_rate: 0.005 | |
| tokenizer: | |
| # Keep vocab 16384 (phase-3 choice) for cross-phase BPB comparability. Retrain | |
| # the BPE on a fresh seeded 200M-character subset of the larger train split. | |
| vocab_size: 16384 | |
| directory: artifacts/phase4/tokenizer | |
| max_train_chars: 200000000 | |
| tokenized: | |
| directory: data/processed/phase4/tokenized | |
| model: | |
| # Unchanged from phase 3: ~345M parameters (26L / 16H / 1024 embd / 1024 ctx). | |
| n_layer: 26 | |
| n_head: 16 | |
| n_embd: 1024 | |
| block_size: 1024 | |
| dropout: 0.0 | |
| training: | |
| device: cuda | |
| dtype: bfloat16 | |
| # Same static shapes that completed phase 3 at 44.3/47.8GB on the L40S. | |
| batch_size: 32 | |
| gradient_accumulation_steps: 8 | |
| # Finalized after flight benchmark (2026-07-20): measured train tokens = | |
| # 2,742,489,368; flight measured 5.812 s/step on the L40S. tokens_per_step = | |
| # 32*8*1024 = 262,144. 27,000 iters => 7.08B processed => ~2.58 epochs, and | |
| # projects 27000*5.812s = ~43.6h / ~$43.2 with margin under the $45 cost cap | |
| # (28,000 would land ~$44.8, too close). Set lr_decay_iters == max_iters. | |
| max_iters: 27000 | |
| learning_rate: 0.0003 | |
| min_lr: 0.00003 | |
| warmup_iters: 300 | |
| lr_decay_iters: 27000 | |
| eval_interval: 500 | |
| eval_iters: 50 | |
| checkpoint_interval: 500 | |
| grad_clip: 1.0 | |
| out_dir: artifacts/phase4 | |
| budget: | |
| # Finalized after tokenize (2026-07-20): measured train 2.742B > floor. The | |
| # binding constraint is the $45 cost cap (= 45.45h at $0.99/h); 28,000 iters | |
| # (~44.7h / ~$44.3) fits it. Hour cap set to 46 for headroom. Do not start | |
| # training if the GPU flight benchmark projects beyond these. | |
| minimum_train_tokens: 2700000000 | |
| maximum_hours: 46 | |
| maximum_usd: 45 | |
| gpu_hourly_usd: 0.99 | |