Awesome

PureMM

PureMM is developed by TencentQQ Multimedia Research Team.

Release

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

Install

pip install -e .

Preparation

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

Evaluation

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

sh eval/eval_mme.sh

MME benchmark

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

<img width="897" alt="企业微信截图_262e4ae1-bfa3-43f5-bcf6-f9abcbf2d533" src="https://github.com/Q-MM/PureMM/assets/14832463/3a1cca00-5365-4242-8352-2b6a49c32ed9">

Acknowledgments

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