backbone: vit_large_patch16_128
backbone_checkpoint_path_or_url: null
backbone_kwargs:
  depth: 24
  embed_dim: 1080
  global_pool: ''
  grad_checkpoint_every: 0
  half_patch_embed: true
  init_values: 1.0
  min_area_for_padding: 0.0
  mlp_ratio: 4
  num_classes: 0
  num_heads: 12
  patch_size:
  - 16
  - 16
  - 8
  pixdim:
  - 0.75
  - 0.75
  - 3.0
  pos_embed: rope
  qkv_bias: true
  rope_kwargs:
    base: 1000.0
  sliding_window_size:
  - 128
  - 128
  - 32
do_snippet_alignment:
  enabled: true
  localization_sigma: 2.0
  localization_tau: 0.1
  localization_weight: 0.004
  projector_mode_image: copy
  projector_mode_text: copy
  rope_mode: physical
  snippet_mode: axis_localization
feature_comb_embed_dim: 1080
feature_comb_num_heads: 12
feature_comb_num_layers: 4
feature_combiner: feat_vit_large
feature_combiner_checkpoint_path_or_url: null
feature_combiner_kwargs:
  global_pool: ''
  init_values: 1.0
  num_classes: 0
  pos_embed: rope
  rope_kwargs:
    base: 100.0
freeze_image_projection: false
freeze_text_projection: false
image_backbone_embed_dim: 1080
image_projection_weights: null
model_name: spectre-large-pretrained
model_settings: null
projection_dim: 512
radfinder_hf_weights: /scratch/local/radiologie/xmed_output/radfinder_hf_export
text_backbone_kwargs:
  grad_checkpoint: false
  lora_alpha: 64.0
  lora_dropout: 0.05
  lora_r: 16
  lora_target_keywords:
  - q_proj
  - k_proj
  - v_proj
  - o_proj
  tokenizer_max_length: 4096
  use_lora: true
text_hidden_size: 1024
text_projection_weights: null
text_tokenizer: Qwen/Qwen3-Embedding-0.6B
text_weights: null
train_img_size:
- 384
- 384
- 128