{
  "schema_version": 2,
  "method": "phase1_qwen35_bbox_hard_mask",
  "weights": "phase1_mask_decoder.safetensors",
  "hidden_size": 4096,
  "decoder_channels": 1024,
  "pixel_shuffle_factor": 32,
  "decoder_architecture": "vision_concat_mlp_dynamic_pixel_shuffle",
  "target": "bbox_hard_proxy_cell_overlap",
  "loss": {
    "type": "bce_plus_dice",
    "bce_weight": 0.5,
    "dice_weight": 0.5
  },
  "parameter_dtype": "float32",
  "checkpoint": false,
  "seed": 42,
  "proxy_mask": "bbox_hard",
  "trainable_parameters": {
    "trainable_parameters": 150943232,
    "lora_parameters": 51265024,
    "mask_decoder_parameters": 99678208
  },
  "merged_model": true,
  "backbone_family": "qwen3_5",
  "backbone_variant": "Qwen3.5-9B",
  "vision_encoder_features": {
    "source": "Qwen3.5 merged vision tower pooler_output",
    "fusion": "residual concat MLP",
    "fusion_input": "LLM final image-token state concatenated with projected post-merge vision token",
    "fusion_mlp": "LayerNorm(8192)-Linear(8192,4096)-GELU-Linear(4096,4096)",
    "initialization": "final fusion linear initialized to zero"
  },
  "base_model": "Qwen/Qwen3.5-9B"
}