VLMEvalKit
VLMEvalKit copied to clipboard
[Model] add kosmos2
- tested on
MMEdataset
observation: It seems that Kosmos2 does not pre-train well with strong knowledge, resulting in poor performance on datasets that require extensive knowledge (testes with MMBench, overall only get 0.013), but it performance is as expected while testing on other dataset which do not require as much knowledge:
-
BLINKdataset -
LLavaBench -
MMMU -
MMStar -
MMVet