model:
  arch: shikra
  version: 'v1'

  # vit encoder
  cache_dir: None
  vit_model: "openai/clip-vit-large-patch14"
  freeze_vit: True

  # finetune config
  freeze_backbone: False
  tune_mm_mlp_adapter: False
  freeze_mm_mlp_adapter: False

  # model config
  mm_vision_select_layer: -2
  model_max_length: 2048

  # data process config
  image_token_len: 256
  mm_use_im_start_end: True

  # training config
  bf16: False
  fp16: True


preprocess:
    vis_processor:
        train:
          name: "clip_image_train"
          proc_type: "openai/clip-vit-large-patch14"
        eval:
          name: "clip_image_eval"
          proc_type: "openai/clip-vit-large-patch14"
    text_processor:
        train:
          name: "blip_caption"
        eval:
          name: "blip_caption"