data:
  manifest_train: <path redacted for release>
  manifest_val: <path redacted for release>
  modality: emg
  weights: temperature:0.5
model:
  channel_embed_dim: 64
  channel_encoding: id
  channel_injection: token
  channel_token_gate_init: 0.0
  depth: 6
  dropout: 0.0
  embed_dim: 384
  ffn: swiglu
  fold_gamma: 0.1
  fold_synthesis: 3
  importance_ratio: 0.6
  mask_ratio: 0.7
  masking_strategy: frequency_guided
  max_level: 3
  mlp_ratio: 4.0
  norm: rmsnorm
  num_heads: 6
  qk_norm: true
  scale_fold: dynamic
  use_separate_channel: true
  wave_init_mode: pad
  wave_kernel_size: 16
  wavelet_names:
  - db2
  - db4
  - db7
  - sym4
  - sym5
objective:
  fold_kl: 0.001
  mask_before_frontend: true
  normalize_spec_target: true
  raw_beta: 0.5
  raw_weight: 0.5
  spec_weight: 0.5
output_dir: <path redacted for release>
run_name: pretrain_emg_c1_moe
seed: 42
train:
  batch_size_by_route:
    A24_2000: 128
    G64_2000: 48
    W16_2000: 192
  clip_grad_norm: 1.0
  epochs: 60
  grad_accumulation_steps: 1
  lr: 0.0003
  max_skipped_steps_per_epoch: 50
  min_lr_ratio: 0.01
  precision: bf16
  skip_step_grad_norm_above: 100.0
  steps_per_epoch: 1000
  val_mask_seed: 1234
  val_max_batches_per_dataset: 50
  warmup_epochs: 2
  weight_decay: 0.05
trainer: emg_c1_moe