hayooucom / visual_model

WeMM

Multi-modal LLM with representation routing. Currently, we have released the model that achieves 2066.66 points on the MME benchmark. The chat version will come later.

News

• 2023-11-20: Inference and mme evaluation code released. Model released.
• 2023-11-13: WeMM scored 2066.66 points on MME full benchmark, which was top1 at the time.
  • Perception score = 1621.66. Ranked the first place.
  • Cognition score = 445. Ranked the third place.
• 2023-10-13: WeMM achieved a perception score of 1551.63 on MME benchmark, which was the state-of-the-art performance at the time.
• 2023-9-13: WeMM scored 68.6 points on MMBench, ranking second place at the time.

Inference

The following command will start an interactive interface that allow users to input image path and prompt and get response from WeMM.

python3 inference.py

MME evaluation

First copy the official MME evaluation files into ./MME_Benchmark_release_version. Then run:

bash mme_eval.sh

MME Benchmark

WeMM achieved 2066.66 points (1621.66 perception score + 445 cognition score), which was top1 on MME full benchmark on 2023-11-13

MMBench

WeMM ranks second place on MMBench on 2023-9-13.