[Paper]
[Website]
WebEyes is a task-level benchmark for evaluating search-based visual reasoning.
It supports three datasets: WebEyes-Ground, WebEyes-Seg, and WebEyes-VQA.
Files
Each task is released as one JSONL file. Each row is one QA instance.
data/search_grounding.jsonl
data/search_segmentation.jsonl
data/search_vqa.jsonl
The Dataset Viewer uses mirrored Parquet files under data/viewer/ so the
image column can be rendered directly on Hugging Face.… See the full description on the dataset page: https://huggingface.co/datasets/yangbokang81/WebEyes.
