PureMM is developed by TencentQQ Multimedia Research Team.

• [2024-01-26]：v1.3 model weight released.
• [2024-01-17]：v1.2 model weight released.
• [2023-12-28]：Model and inference code released.

pip install -e .

• The latest version of model weight should download from huggingface: https://huggingface.co/Q-MM/PureMM/tree/main

Before evaluation, please download the MME benchmark files MME into the mme_bench directory.
Then run the cmd:

sh eval/eval_mme.sh

PureMM achieved 1686.52 points on perception score, which was top1 on MME benchmark on 2024-01-21.

• LLaVA: the codebase we built upon. Thanks for their wonderful work.
• Vicuna: the amazing open-sourced large language model!