search

microsoft / LLaVA-Med

Large Language-and-Vision Assistant for Biomedicine, built towards multimodal GPT-4 level capabilities.
Other
1.29k stars 148 forks source link

Projector weight for model #34

Open TrungQuocNguyen opened 7 months ago

TrungQuocNguyen commented 7 months ago

Dear authors,

Thank you very much for this amazing paper. I tried to reproduce your results in table 4 by using the finetuned weights of VQA-RAD to evaluate on the downstream dataset, but it seems to me that only the finetuned weights of LLM (i.e LLaVA-Med) is released, while the projector weight is missing. I also checked the repo of LLaVA and noticed that they did release the projector weights as well. Without projector weights, we cannot reproduce the results in your paper. I would really appreciate if you can provide them so we can try with our own experiments. Thank you !

bangawayoo commented 4 months ago

Hi. In the delta weights provided by the authors, there seem to be what should be the projector's parameters (model.mm_projector.weights and model.mm_projector.bias). Are they not the projector weights?

Thanks.