experiment:
  name: p4_tiny_backbone
  seed: 1
  mode: backbone_only
  model:
    name: tiny_p4
    vocab_size: 128
    d_model: 192
    n_layers: 4
    n_heads: 6
    n_kv_heads: 2
    ffn_width: 512
    max_seq_len: 320
    w: 64
    ffn_w: 192
    n_slots: 16
    n_editor_heads: 2
    r_max: 8
  init:
    freeze_backbone: false
  data:
    kind: synthetic
    tokenizer: char
    splits_dir: data/splits/synthetic_v1
    loss_span: all
    eval_splits:
    - dev_iid
    - dev_template
    - dev_depth
    - dev_compose
  train:
    batch_size: 64
    seq_len: 320
    steps: 6000
    lr: 0.001
    warmup: 200
    weight_decay: 0.1
    grad_clip: 1.0
    log_every: 50
    eval_every: 500
    eval_batches: 10
    ckpt_every: 2000
  em_max_items: 600
model:
  name: tiny_p4
  vocab_size: 128
  d_model: 192
  n_layers: 4
  n_heads: 6
  n_kv_heads: 2
  ffn_width: 512
  max_seq_len: 320
  rope_theta: 10000.0
  norm_eps: 1.0e-05
  tie_embeddings: true
  init_std: 0.02
  use_workspace: false
  w: 64
  ffn_w: 192
  n_slots: 16
  n_editor_blocks: 2
  n_editor_heads: 2
  r_max: 8
  eta_max: 0.25
  clip_c: 1.0
  clip_eps: 1.0e-06
  readout_zero_init: true
  controller_hidden: 0
  lambda_cost: 0.0
  kappa_risk: 0.0
  dense_writes: false
resolved_train:
  batch_size: 64
  seq_len: 320
  steps: 6000
  lr: 0.001
  warmup: 200
  weight_decay: 0.1
  grad_clip: 1.0
  log_every: 50
  eval_every: 500
  eval_batches: 10
  ckpt_every: 2000