datasets:

  llava_pretrain:
    data_type: image
    sample_ratio: 1

  # internvid:
  #   data_type: frames
  #   sample_ratio: 10
  
  gpt4v_public:
    data_type: frames
    sample_ratio: 1
    task_types: ['summary']
    fps: 1.0
    conv_type: single
    train_data_path: /mnt/bn/algo-masp-nas-2/xiangchen/data/shared_gpt4v_data/data_500k_filtered.json