name: p4_diag_b8only_w128
seed: 1
mode: prisme
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:
  backbone_checkpoint: experiments/runs/p4_tiny_backbone_seed01_20261029232626/checkpoint
  freeze_backbone: true
  auto_readout_scale: true
data:
  kind: synthetic
  tokenizer: char
  splits_dir: data/splits/synthetic_v1
  loss_span: answer
  eval_splits:
  - dev_iid
  - dev_template
  - dev_depth
  - dev_compose
budgets:
- 8
budget_probs:
- 1.0
budget_per: position
train:
  batch_size: 32
  seq_len: 320
  steps: 3000
  lr: 0.001
  warmup: 100
  weight_decay: 0.1
  grad_clip: 1.0
  log_every: 50
  eval_every: 500
  eval_batches: 12
  ckpt_every: 1500
em_budgets:
- 0
- 2
- 8
em_max_items: 400
model_overrides:
  readout_zero_init: false
  readout_init_std: 0.2
  w: 128
  ffn_w: 384
  n_editor_heads: 4