processing/phase2.yaml

extends: configs/base.yaml
phase: phase2
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
  ndl_selection_manifest: data/processed/ndl/selection.jsonl
  ndl_download_manifest: data/processed/ndl/downloads.jsonl
  ndl_raw_dir: data/raw/ndl/books
  ndl_clean_dir: data/clean/ndl
  ndl_clean_manifest: data/processed/ndl/clean.jsonl
  split_manifest: data/processed/phase2/splits.jsonl
  ndl_max_abnormal_ngram_rate: 0.38
  ndl_max_latin_rate: 0.005
tokenizer:
  directory: artifacts/phase2/tokenizer
tokenized:
  directory: data/processed/phase2/tokenized
model:
  # 16 layers compensate for the much smaller Japanese vocabulary embedding;
  # 12 layers with an 8k vocabulary would only be about 91M parameters.
  n_layer: 16
  n_head: 12
  n_embd: 768
  block_size: 1024
  dropout: 0.0
training:
  device: cuda
  dtype: bfloat16
  batch_size: 4
  gradient_accumulation_steps: 32
  max_iters: 1600
  learning_rate: 0.0006
  min_lr: 0.00006
  warmup_iters: 100
  lr_decay_iters: 1600
  eval_interval: 100
  eval_iters: 50
  checkpoint_interval: 100
  grad_clip: 1.0
  out_dir: artifacts/phase2
budget:
  minimum_train_tokens: 200000000
  maximum_hours: 20
  maximum_usd: 6
  gpu_hourly_usd: 0.27