datasets:

  llava_pretrain:
    data_type: image
    sample_ratio: 1

  # gpt4v_public:
  #   data_type: frames
  #   sample_ratio: 2
  #   task_types: ['summary']
  #   fps: 1.0
  #   conv_type: single

  # lk_image:
  #   data_type: image

  # lk_video:
  #   data_type: frames
  #   conv_type: multi
  #   fps: 1.0

  gpt4v_internal:
    data_type: frames
    fps: 2.0
    sample_ratio: 1
    conv_type: multi
    task_types: ['qa_pairs']