{
  "architectures": [
    "LaMEMultimodal"
  ],
  "ctl_loss_weight": 1.0,
  "decode_loss_weight": 1.0,
  "decode_target": "answer",
  "decoder_gradient_checkpointing": false,
  "decoder_max_think_tokens": 1024,
  "decoder_model_path": "Qwen/Qwen3-0.6B",
  "decoder_projector_hidden_dim": 2048,
  "decoder_projector_num_layers": 1,
  "diversity_loss_weight": 0.1,
  "dtype": "bfloat16",
  "embedding_token_id": 151665,
  "enable_decode": true,
  "freeze_backbone": false,
  "freeze_decoder": false,
  "lora_decoder_enable": false,
  "model_type": "openlame",
  "normalize": true,
  "num_new_special_tokens": 9,
  "num_reason_tokens": 8,
  "pooling_strategy": "special_token",
  "projection_activation": "gelu",
  "projection_dim": 3584,
  "projection_dropout": 0.0,
  "projection_hidden_dim": null,
  "projection_num_layers": 1,
  "reason_ctl_loss_weight": 1.0,
  "reasoning_embedding_type": "mean_pooling",
  "split_latent": true,
  "training_stage": 2,
  "transformers_version": "4.57.0",
  "trust_remote_code": true,
  "use_flash_attn": true,
  "use_projection": false,
  "vlm_name": "Qwen/Qwen2-VL-2B-Instruct",
  "vlm_torch_dtype": "bfloat16",
  "vocab_size": 151666
}