Vision-CAIR / MiniGPT-4

Open-sourced codes for MiniGPT-4 and MiniGPT-v2 (https://minigpt-4.github.io, https://minigpt-v2.github.io/)
https://minigpt-4.github.io
BSD 3-Clause "New" or "Revised" License
25.29k stars 2.9k forks source link

What about with v1.5 - LLAMA2 13B OR 13B-16k ? #312

Open tuyra opened 1 year ago

tuyra commented 1 year ago

Hi,

Has anyone tried the new version of Vicuna weights? Does it work?

Thanks.

Sypherd commented 1 year ago

You would have to train the projection layer from scratch as well as perform the second stage finetuning again to work with Llama 2 weights. There have also been reports of people switching to different models and the image input still working but that is not reliable.