This isn't Alpaca 65B
You're just git cloning Llama and then quantizing it with Alpaca.cpp. the quantization script is basically the same for both alpaca and llama. And I don't see any part where you merge the Lora weights with the llama model.
Yes, I did not know that where is alpaca.cpp 65B model weight, that is llama 65B q_4 weight .
You are right.
it run using alpaca.cpp
Yes, I did not know that where is alpaca.cpp 65B model weight, that is llama 65B q_4 weight
it run using alpaca.cpp
Then what's the difference between this and the original LLaMa 65B?
So here is my attempt to explain Alpaca vs LLaMA, and why it's hard to find Alpaca.
So LLaMA predicts the next word, it is not great at following instructions or chatting. If you chat with it, you will probably be disappointed.
Once we teach LLaMA to follow instructions and chat it becomes Alpaca, which is more comparable to ChatGPT. Fun!
There are two common ways to get to Alpaca:
- fine tuning LLaMA which results is a new model
- LoRA which results in a small add-on to the base LLaMA model (they can be merged if needed) usually people use the huggingface/peft library . It needs to be combined with the LLaMA base model to work. On one hand the lora addon is small to upload and is good for training on smaller computers. But on the other it can be less compatible.
Most people are using LoRA so that's why it can be hard to find a full and compatible Alpaca model.
Don't worry this will all be clearer in the next few weeks as people make more formats available. And the Alpaca dataset is being cleaned which is actually leading to differen't version of Alpaca, depending on if they use the original dataset, the old cleaned dataset, or the latest cleaned alpaca dataset.
So here is my attempt to explain Alpaca vs LLaMA, and why it's hard to find Alpaca.
I know what the difference between Alpaca and LLaMA is. But this one is the exact same model as LLaMA. By "this" I mean the model in this repo, and by "the model in this repo", I mean LLaMA. So what I'm really asking is "what's the difference between LLaMA and LLaMA? And the answer is nothing. The model in this repo is LLaMA 65B. It isn't Alpaca 65B at all.
Also my question was a half-rhetorical question.
Oh right, I got you.
Yeah xfh might want to change the name of the repo.