I’m trying to fine tune Llama 3.1 manually (without using SFTTrainers) to combine it with my image classification model. I need to convert the output I get from the model training which is a bunch of logits to text to check if Llama learned to mention a few facts I have extracted from the image, if not then penalize it. without this, the model tries to only generate texts similar to the facts I give it.
1 Like