How this has been trained?
Just curious, does it produce better output than your Alpaca-13B Native 4-bit?
I believe this is a lora
is it quantized? 4-bit? 8-bit?
just lookkkkkkkkkkkkkkkkkkkkkkkkkkkkkkkkkkkkkkkkkkkkkkkkkkkkkkkkkkkkkkkkkkkkkkkkkkkkkkkkkkkkkkkkkkkkkkkkkkkkkkkkkkkkkkkkkkkkkkkkkkkkkkkkkkkkkkkkkkkkkkkkkkkkkkkkkkkkkkkkkkkk,.
its a rank 8 lora
your Alpaca-13B
but you can see in the config that it's 7b??
its currently checkpoint 1000, not done!
@Yoshiii Well sorry if I'm trying to learn something. While we're at it help me a sec to clear up my mind, some random questions in no particular order: does passing the --bf16 flag change anything for regular usage (no training)? If I got a LLaMA model with Alpaca's LoRA baked in, should I still include the very same LoRA in the model's folder or the baked one will do the trick by itself? Then, it came as a single safetensor file, I downloaded the proper (I think) HF LLaMA weights and put the model in that folder: now, as far as I understand I have an Alpaca model and the weights are for LLaMA, what am I missing here? Also can/should I delete the original HF safetensor (weight?) from the folder or is it still needed? And finally, is there anything else that I should do/edit in the goddamn folder? Thank you