Question on the two variants
You provided two different fine tunes for 13B, both are very differently named (one pretrained vicuna, and this one)
The readme lists less training data for the vicuna upload, the configuration shows "square" as aspect ratio, does that mean the image is cropped to square instead of padding in this case ?
What is the purpose of the other finetune ?
I'm just testing both, the vicuna one appears to hallucinate a tiny bit more than the previous 7B variant and it does not follow instructions as good as I was used to.
You provided two different fine tunes for 13B, both are very differently named (one pretrained vicuna, and this one)
The readme lists less training data for the vicuna upload, the configuration shows "square" as aspect ratio, does that mean the image is cropped to square instead of padding in this case ?What is the purpose of the other finetune ?
I'm just testing both, the vicuna one appears to hallucinate a tiny bit more than the previous 7B variant and it does not follow instructions as good as I was used to.
Sorry for the confusion. The vicuna one is LLM only after the pretain stage.
Thanks a lot, that explains it.
I tested the other one and it does not have the issues anymore and is better in listing fine details than the 7B one.
However - hallucinations increased also
Thanks a lot, that explains it.
I tested the other one and it does not have the issues anymore and is better in listing fine details than the 7B one.
However - hallucinations increased also
Thanks for your feedback! We will make some efforts to decrease the hallucinations in future work.