SPB Git

spb/wp31_uqo

Public
1.8 KB · 111 lines yaml
1experiment:2  seed: 13  model:4    name: tiny_p4v25    vocab_size: 1286    d_model: 2567    n_layers: 68    n_heads: 89    n_kv_heads: 410    ffn_width: 76811    max_seq_len: 32012    w: 6413    ffn_w: 19214    n_slots: 815    n_editor_heads: 216    r_max: 817    readout_zero_init: false18    readout_init_std: 0.219  init:20    backbone_checkpoint: experiments/runs/p4_tiny_backbone_v2_seed01_20261029233624/checkpoint21    freeze_backbone: true22    auto_readout_scale: true23  data:24    kind: synthetic25    tokenizer: char26    splits_dir: data/splits/synthetic_v127    loss_span: answer28    eval_splits:29    - dev_iid30    - dev_template31    - dev_depth32    - dev_compose33  budgets:34  - 035  - 136  - 237  - 438  - 839  budget_probs:40  - 0.241  - 0.242  - 0.243  - 0.244  - 0.245  budget_per: position46  train:47    batch_size: 3248    seq_len: 32049    steps: 300050    lr: 0.00151    warmup: 10052    weight_decay: 0.153    grad_clip: 1.054    log_every: 5055    eval_every: 50056    eval_batches: 1257    ckpt_every: 300058  em_budgets:59  - 060  - 161  - 262  - 863  em_max_items: 100064  name: p4v2_abl_M865  mode: prisme66model:67  name: tiny_p4v268  vocab_size: 12869  d_model: 25670  n_layers: 671  n_heads: 872  n_kv_heads: 473  ffn_width: 76874  max_seq_len: 32075  rope_theta: 10000.076  norm_eps: 1.0e-0577  tie_embeddings: true78  init_std: 0.0279  use_workspace: true80  w: 6481  ffn_w: 19282  n_slots: 883  n_editor_blocks: 284  n_editor_heads: 285  pair_scheme: adjacent86  r_max: 887  eta_max: 0.2588  clip_c: 1.089  clip_eps: 1.0e-0690  readout_zero_init: false91  readout_init_std: 0.292  readout_scale: 1.093  controller_hidden: 094  ctrl_input: z95  ctrl_use_entropy: false96  lambda_cost: 0.097  kappa_risk: 0.098  dense_writes: false99resolved_train:100  batch_size: 32101  seq_len: 320102  steps: 3000103  lr: 0.001104  warmup: 100105  weight_decay: 0.1106  grad_clip: 1.0107  log_every: 50108  eval_every: 500109  eval_batches: 12110  ckpt_every: 3000111
Files · main
Menu

Theme