The Circular Evaluation Strategy && LLM-based Choice Extractors
yyy809643573 opened this issue · comments
yyy809643573 commented
hi, i am using the MMBench to evaluate my VLM, but i find there are some ploblems in MMBench
- I see that in https://github.com/open-compass/opencompass/blob/main/configs/multimodal/minigpt_4/README.md, the minigpt4 does not achieve the The Circular Evaluation Strategy and LLM-based Choice Extractors by using chatgpt,Are The Circular Evaluation Strategy and LLM-based Choice Extractors needed to achieve by ourselves?
- the question number of MMBench-Dev(En) is not 1164 in describtion of https://github.com/open-compass/MMBench
Looking forward to your reply,Thank you!
Haodong Duan commented
- CircularEval and LLM-based Choice Extractors are implemented in eval_mmbench.py, which is not included in the inference part.
- The question number in the MMBench-Dev tsv is more than 1164, since it also includes CircularEval cases with choice shifted.