Laser-ScanPath

GPT-4o scan-path annotations for LASER training. 269,773 samples in gpt4o_train_full_filter.json.

Each sample has an image field with paths like viscot/gqa/2403340.jpg. The images are not included here — download them from Visual-CoT and arrange them so the relative paths resolve:

images/
└── viscot/
    ├── flickr30k/
    ├── gqa/
    ├── openimages/
    ├── v7w/
    ├── cub/
    └── vsr/

Only these 6 subsets of Visual-CoT are used.

Using with the LASER training script

meta_data_reasoning_chain_full_filter.json is a small wrapper that the LASER trainer takes as --data_path. Edit image_folder to point at your local images root:

[{
    "ds_name": "laser_scanpath",
    "data_path": "gpt4o_train_full_filter.json",
    "image_folder": "/PATH/TO/YOUR/images/",
    "ds_type": "reasoning_chain"
}]