name: baseline-lm-only
records: results/records.jsonl
model:
  n_layer: 6
  n_head: 6
  n_embd: 384
  block_size: 512
  dropout: 0.0
training:
  seed: 0
  steps: 6000
  batch_size: 16
  lr: 0.0003
  warmup: 200
  log_every: 100
  device: cuda
  tokenizer: data/tokenizer.json
  episodes_dir: data/episodes
  mixture:
  - env: lm/fineweb-edu
    kind: lm_tokens
    weight: 1.0
    split: train
    weighting: ml
    kwargs:
      corpus_dir: data/corpus-fineweb-edu-40000000