experiment:
  seed: 1
  model: configs/prisme_103m.yaml
  model_overrides:
    readout_zero_init: false
    readout_init_std: 0.2
    ctrl_input: h
    ctrl_use_entropy: true
  init:
    backbone_checkpoint: experiments/runs/pretrain103m_seed01_20261029232706/checkpoint
    freeze_backbone: true
    auto_readout_scale: true
  data:
    kind: tokens
    train_bin: data/processed/fineweb_edu_300m/train.bin
    val_bin: data/processed/fineweb_edu_300m/val.bin
  budgets:
  - 0
  - 1
  - 2
  - 4
  - 8
  budget_probs:
  - 0.2
  - 0.2
  - 0.2
  - 0.2
  - 0.2
  budget_per: position
  train:
    batch_size: 16
    seq_len: 512
    steps: 3000
    lr: 0.0003
    warmup: 100
    weight_decay: 0.1
    grad_clip: 1.0
    log_every: 20
    eval_every: 500
    eval_batches: 10
    ckpt_every: 1000
  final_exact_match: false
  name: editor_pilot_103m_lr3e4
  mode: prisme
model:
  name: prisme_103m
  vocab_size: 32768
  d_model: 768
  n_layers: 12
  n_heads: 12
  n_kv_heads: 3
  ffn_width: 2048
  max_seq_len: 4096
  rope_theta: 10000.0
  norm_eps: 1.0e-05
  tie_embeddings: true
  init_std: 0.02
  use_workspace: true
  w: 256
  ffn_w: 768
  n_slots: 16
  n_editor_blocks: 2
  n_editor_heads: 4
  r_max: 8
  eta_max: 0.25
  clip_c: 1.0
  clip_eps: 1.0e-06
  readout_zero_init: false
  readout_init_std: 0.2
  readout_scale: 1.0
  controller_hidden: 0
  ctrl_input: h
  ctrl_use_entropy: true
  lambda_cost: 0.0
  kappa_risk: 0.0
  dense_writes: false
resolved_train:
  batch_size: 16
  seq_len: 512
  steps: 3000
  lr: 0.0003
  warmup: 100
  weight_decay: 0.1
  grad_clip: 1.0
  log_every: 20
  eval_every: 500
  eval_batches: 10
  ckpt_every: 1000