# Standalone SL recipe; run from the repository root after downloading data.
factory: embodichain_models.dynamics.nerd.sl_recipe:build_supervised_trainer
model:
  input_dims:
    states_embedding: 29
    contact_normals: 75
    contact_points_1: 75
    contact_depths: 25
    joint_f: 8
    gravity_dir: 3
  output_dim: 29
  n_embd: 384
  n_layer: 6
  n_head: 12
  block_size: 32
  history_length: 10
  head_hidden_dims:
  - 64
  bias: false
  dropout: 0.0
  normalize_input: true
  normalize_output: true
  attention_backend: math
data:
  path: outputs/datasets/neural_dynamics/nerd/ant/v1
  # Keep full-corpus training on bounded reads unless a fitting budget is set.
  resident_cache_bytes: 0
objective:
  include_prefix: false
  numerical_profile: reference
optimizer:
  learning_rate: 0.001
  final_learning_rate: 0.0001
  lr_update_interval: 500
  lr_offset: 1
  max_grad_norm: 1.0
trainer:
  total_updates: 1000
  batch_size: 512
  device: cuda:0
  seed: 0
  output_dir: outputs/sl/nerd/ant
  checkpoint_interval: 500
  validation_interval: 500
evaluation:
  max_batches: 100