config.json

{
  "_name_or_path": "RAS-13B-General",
  "architectures": [
    "RASLlamaForCausalLM"
  ],
  "attention_bias": false,
  "attention_dropout": 0.0,
  "bos_token_id": 1,
  "eos_token_id": 2,
  "exclusive_mask_decoding": true,
  "freeze_mask_classifier": false,
  "freeze_mask_projector": false,
  "freeze_mask_reprojector": false,
  "freeze_vision_projector": false,
  "freeze_vision_tower": true,
  "hidden_act": "silu",
  "hidden_size": 5120,
  "image_aspect_ratio": "pad",
  "initializer_range": 0.02,
  "intermediate_size": 13824,
  "mask_decoding_binary_threshold": 0.5,
  "mask_loss_binary": false,
  "mask_loss_binary_pos_weight": 5.0,
  "mask_loss_direct_binary": true,
  "mask_loss_set_pred": true,
  "mask_loss_type": "mse+infonce",
  "mask_projector_input_features": [
    "clip_-2",
    "siglip",
    "open_clip",
    "dinov2",
    "sincos_pe"
  ],
  "mask_projector_type": "mlp2x_gelu",
  "mask_reprojector_type": "mlp2x_gelu",
  "max_length": 4096,
  "max_position_embeddings": 4096,
  "mm_use_im_patch_token": false,
  "mm_use_im_start_end": false,
  "model_type": "ras_llama",
  "num_attention_heads": 40,
  "num_hidden_layers": 40,
  "num_key_value_heads": 40,
  "pad_token_id": 0,
  "pretrain_projectors": false,
  "pretraining_tp": 1,
  "rms_norm_eps": 1e-05,
  "rope_scaling": null,
  "rope_theta": 10000.0,
  "tie_word_embeddings": false,
  "torch_dtype": "bfloat16",
  "transformers_version": "4.37.2",
  "use_cache": true,
  "version": "v1",
  "vision_backbone_resolutions": [
    336,
    378,
    1024,
    518,
    448
  ],
  "vision_backbones": [
    "openai/clip-vit-large-patch14-336",
    "google/siglip-so400m-patch14-384",
    "convnext_xxlarge",
    "facebook/dinov2-large",
    "sincos_pe"
  ],
  "vision_hidden_size": [
    1024,
    1152,
    3072,
    1024,
    1024
  ],
  "vision_output_features": [
    "clip_-2",
    "siglip",
    "open_clip",
    "dinov2",
    "sincos_pe"
  ],
  "vision_pe_hidden_size": 1024,
  "vision_pe_patches_per_side": 32,
  "vision_pe_scale": 1.0,
  "vision_processor": "openai/clip-vit-large-patch14-336",
  "vision_projector_input_features": [
    "clip_-2"
  ],
  "vision_projector_type": "mlp2x_gelu",
  "vocab_size": 32004
}