training:
  model_name: jsfakes_tonnetz_llamathree
  batch_size: 48
  lr: 0.001
  lr_warmup_steps: 81
  lr_decay_until_steps: 812
  lr_decay_factor: 0.001
  weight_decay: 0.1
  amp_precision: bfloat16
  weight_precision: float32
  enable_mixed_precision: true
  num_epochs: 5
  output_dir: output/jsfakes_tonnetz_llamathree
  save_every_step: 500
  log_every_step: 10
  wandb_project: tonnetz
  torch_compile: false
model:
  type: llamathree
  context_length: 512
  emb_dim: 64
  n_heads: 4
  n_layers: 4
  hidden_dim: 128
  hidden_activation: silu
  n_kv_groups: 1
  rope_base: 50000
  rope_freq: null
  dtype: float32
  vocab_size: 706
dataset:
  hugging_face_id: TristanBehrens/jsfakes_tonnetz
tokenizer:
  type: whitespace
  fill_token: '[EOS]'