Why not the new prompt format?
Hi, I am wondering why you didn't use the new prompt format, given that the model already was heavily finetuned on that?
This finetune is based on the Llama-2 base model, not the chat model. The base model does not have a prompt format as it is not an instruct tuned model. If it was based on the chat model then it would indeed make more sense to stick to that format. But generally speaking it is always best to finetune base models, especially in this case where the chat model has already been very heavily tuned.
Also since I'm interested in doing like-for-like comparisons I wanted the model training to be as close to the original as possible. Which is why I used the exact script that was used to train the original Guanaco model with basically 0 modifications. I'm also relatively new when it comes to finetuning so I wanted to stick with something proven, to try to decrease the chances that I screw something up.
Oh, I totally missed that this was trained on the base model! I completely get it, thank you for the comment!
Did you come across evaluating performance? :)