name: state_halt
checkpoint: null
device: cuda

kwargs:
  cell_set_len: 512
  blur: 0.05
  hidden_dim: 768
  loss: energy
  confidence_token: True
  n_encoder_layers: 1
  n_decoder_layers: 1
  predict_residual: True
  softplus: True
  freeze_pert_backbone: False
  transformer_decoder: False
  finetune_vci_decoder: False
  residual_decoder: False
  batch_encoder: False
  use_batch_token: False
  mask_attn: False
  # --- adaptive-depth halting (joint calibration, user §4.4/§4.5) ---
  n_refine_rounds: 6
  halt_tau: 0.05
  halt_alpha: 0.5
  halt_beta: 0.1
  halt_gamma: 0.01
  halt_warmup_steps: 2000
  halt_magnitude_free: true
  confidence_weight: 0.01
  confidence_target_scale: 10.0

  use_effect_gating_token: False
  distributional_loss: energy
  init_from: null
  mmd_num_chunks: 1
  randomize_mmd_chunks: false
  llm_name: null

  transformer_backbone_key: llama
  transformer_backbone_kwargs:
      bidirectional_attention: true
      max_position_embeddings: ${model.kwargs.cell_set_len}
      hidden_size: ${model.kwargs.hidden_dim}
      intermediate_size: 3072
      num_hidden_layers: 8
      num_attention_heads: 12
      num_key_value_heads: 12
      head_dim: 64
      use_cache: false
      attention_dropout: 0.0
      hidden_dropout: 0.0
      layer_norm_eps: 1e-6
      pad_token_id: 0
      bos_token_id: 1
      eos_token_id: 2
      tie_word_embeddings: false
      rotary_dim: 0
      use_rotary_embeddings: false

  lora:
      enable: false
      r: 16
      alpha: 32
      dropout: 0.05
      bias: none
      target: auto
      adapt_mlp: false
      task_type: FEATURE_EXTRACTION
      merge_on_eval: false