SPB Git

spb/wp31_uqo

Public
1.8 KB · 107 lines yaml
1experiment:2  name: mini_editor3  seed: 14  mode: prisme5  model:6    name: mini7    vocab_size: 1288    d_model: 969    n_layers: 210    n_heads: 411    n_kv_heads: 212    ffn_width: 25613    max_seq_len: 32014    w: 3215    ffn_w: 9616    n_slots: 1617    n_editor_heads: 218    r_max: 819    readout_zero_init: false20    readout_init_std: 0.221    ctrl_input: h22    ctrl_use_entropy: true23  init:24    backbone_checkpoint: experiments/runs/mini_backbone_seed01_20261030005114/checkpoint25    freeze_backbone: true26    auto_readout_scale: true27  data:28    kind: synthetic29    tokenizer: char30    splits_dir: data/splits/synthetic_v131    loss_span: answer32    eval_splits:33    - dev_iid34  train:35    batch_size: 1636    seq_len: 32037    steps: 12038    lr: 0.00139    warmup: 1040    weight_decay: 0.141    grad_clip: 1.042    log_every: 5043    eval_every: 12044    eval_batches: 245    ckpt_every: 12046  em_max_items: 3247  budgets:48  - 049  - 150  - 251  - 452  - 853  budget_probs:54  - 0.255  - 0.256  - 0.257  - 0.258  - 0.259  budget_per: position60  em_budgets:61  - 062  - 263model:64  name: mini65  vocab_size: 12866  d_model: 9667  n_layers: 268  n_heads: 469  n_kv_heads: 270  ffn_width: 25671  max_seq_len: 32072  rope_theta: 10000.073  norm_eps: 1.0e-0574  tie_embeddings: true75  init_std: 0.0276  use_workspace: true77  w: 3278  ffn_w: 9679  n_slots: 1680  n_editor_blocks: 281  n_editor_heads: 282  r_max: 883  eta_max: 0.2584  clip_c: 1.085  clip_eps: 1.0e-0686  readout_zero_init: false87  readout_init_std: 0.288  readout_scale: 1.089  controller_hidden: 090  ctrl_input: h91  ctrl_use_entropy: true92  lambda_cost: 0.093  kappa_risk: 0.094  dense_writes: false95resolved_train:96  batch_size: 1697  seq_len: 32098  steps: 12099  lr: 0.001100  warmup: 10101  weight_decay: 0.1102  grad_clip: 1.0103  log_every: 50104  eval_every: 120105  eval_batches: 2106  ckpt_every: 120107
Files · main
Menu

Theme