# Laptop smoke-test model (~7.4M params, ~20M tokens on CPU).
name: tiny
model:
  d_model: 256
  n_layers: 4
  n_heads: 4
  ffn_hidden: 704
  context_len: 512
train:
  peak_lr: 1.0e-3
  total_steps: 2440
  micro_batch_size: 8
  grad_accum: 4
  precision: fp32
  data_dir: data/processed
  mix: {astro: 0.30, arxiv: 0.15, textbook: 0.05, general: 0.50}
  max_epochs: {textbook: 3, astro: 2}