Abst
OwlEval
Experimental results show that their model is superior
Introduction
They propose mPLUG-Owl, a novel training paradigm
They carefully construct an instruction evaluation set, called as OwlEval.
Related works
Large language model
BERT
GPT
T5
Multi-modai large language models
Visual ChatGPT
MM-REACT
HuggingGPT
unified model
CLIP
BLIP,BLIP2
mPLUG-Owl not only aligns the representation between the vision and language foundation model.
Mainly 2 steps
Step1
Multimodal pertaining
Step2
Joint Instruction tuning
OwlEval is constructed about 80 questions and 50 images.
感想
ベンチマークにするにしてはデータ少ない気が
Top comments (0)