extends: configs/base.yaml
phase: phase1
data:
  ndl_clean_manifest: data/processed/ndl/clean.jsonl
  split_manifest: data/processed/phase1/splits.jsonl
  phase1_target_clean_chars: 32000000
tokenizer:
  directory: artifacts/phase1/tokenizer
tokenized:
  directory: data/processed/phase1/tokenized
budget:
  maximum_hours: 1
  maximum_usd: 0.3
  gpu_hourly_usd: 0.27
model:
  n_layer: 8
  n_head: 8
  n_embd: 512
  block_size: 512
  dropout: 0.0
training:
  device: cuda
  dtype: bfloat16
  batch_size: 16
  gradient_accumulation_steps: 8
  max_iters: 3200
  learning_rate: 0.0006
  min_lr: 0.00006
  warmup_iters: 100
  lr_decay_iters: 3200
  eval_interval: 200
  eval_iters: 50
  checkpoint_interval: 200
  grad_clip: 1.0
  out_dir: artifacts/phase1