training:
  model_name: bach_garland_xlstm
  batch_size: 4
  lr: 0.001
  lr_warmup_steps: 5000
  lr_decay_until_steps: 50000
  lr_decay_factor: 0.001
  weight_decay: 0.1
  amp_precision: bfloat16
  weight_precision: float32
  enable_mixed_precision: true
  num_epochs: 4
  output_dir: output/bach_garland_xlstm
  save_every_step: 500
  log_every_step: 10
  wandb_project: bach_garland_xlstm
  torch_compile: false
model:
  num_blocks: 4
  embedding_dim: 64
  mlstm_block:
    mlstm:
      num_heads: 4
  slstm_block:
    slstm:
      num_heads: 4
  slstm_at:
  - 2
  context_length: 4096
  vocab_size: 178
dataset:
  hugging_face_id: TristanBehrens/bach_garland_2024-100K
tokenizer:
  type: whitespace
  fill_token: '[EOS]'