BEAR-benchmark

简介

BEAR Benchmark Data + runnable evaluation code for the BEAR benchmark. Each task folder holds its data (*.json + images/videos) and a run.sh. The shared runners and util/ live at the repository root. You run in two steps: Inference — a VLM answers every question → produces a final_*.json. Scoring — eval.py uses a GPT judge (for multiple-choice) or geometry (for pointing/bbox) to grade those replies and print the final accuracy. 1. Setup pip install -r… See the full description on the dataset page: https://huggingface.co/datasets/yqi19/BEAR-benchmark.