MMR-Bench V2 is a normalized dataset for offline multimodal model-routing
research. It joins canonical multimodal benchmark instances with previously
collected model outcomes so that routing methods can learn or evaluate which
vision-language model to select for each input.
Release v2.0.1 contains:
25,504 instances from 18 benchmarks
1,091,908 model-result records covering
44 models
1,087,897 successful results and 4,011 explicit… See the full description on the dataset page: https://huggingface.co/datasets/gh0stHunter/MMR-Bench-V2.
