kind: nla_model
schema_version: 2
d_model: 1536
tokens:
  injection_char: <|image_pad|>
  injection_token_id: 151655
  injection_left_neighbor_id: 220
  injection_right_neighbor_id: 151645
  critic_suffix_ids: null
prompt_templates:
  av: '<|im_start|>system

    You interpret neural-network activations. Given one activation vector, name in a single sentence the
    concept, entity, topic, or syntactic role it encodes. Be concrete; do not hedge or add preamble.<|im_end|>

    <|im_start|>user

    Activation: {injection_char}<|im_end|>

    <|im_start|>assistant

    '
  ar: '{explanation}'
created_by: nla (dormantx/NLA_Qwen2.5_1.5B pipeline)
role_aliases:
  verbalizer: actor
  recon: critic
layer: 18
extraction_layer_index: 18
base_model: Qwen/Qwen2.5-1.5B-Instruct
role: ar
stage: sl
extraction:
  injection_scale: null
  mse_scale: 1.0
critic:
  extraction_layer_index: 18
training:
  lr: 5.0e-05
  loss_type: mse_unit_norm
  global_batch_size: 32
  num_layers: 18
value_head:
  file: value_head.safetensors
  dtype: float32
  weight_shape:
  - 1536
  - 1536
  bias: true