Is there a way to fine tune the llama.7B model?

[TNTTheLagger]

I want to try and fine-tune this model to see if I can make it into a sort of chatbot.
I have plenty of chat data in json files but I don't know how exactly would I fine-tune the llama model.

Does anyone have any references or tutorials like videos or GitHub repos on this subject?

[DrakoHyena]

It wouldnt be feasible to fine tune it youself, you would need to have quite a few powerful machines. The alpaca model that comes installable is already fine tuned to be a chatbot of sorts

[TNTTheLagger]

Are you referring to the alpaca.7B?

[DrakoHyena]

Yes I am

[TNTTheLagger]

I don't know it acts a bit weird also it tends to just repeat the same sentence or sentences over and over and over without stopping no matter when i set the output token count to

[rafaelpleite]

Try to write in the prompt as below

Below is an instruction that describes a task. Write a response that appropriately completes the request.
Question: [Your question]

like

Below is an instruction that describes a task. Write a response that appropriately completes the request.
Question: What is an alpaca? How is it different from a llama?

[Juleffel]

https://github.com/tloen/alpaca-lora with a few changes seem to be a good option. (Just by changing the generate_prompt to something more chat conversation oriented than instructions oriented).

https://github.com/deep-diver/Alpaca-LoRA-Serve is also a good option: he is using the alpaca-lorca model with a prompt generated to give the instruction "give an answer to the following conversation remembering the context", but it doesn't seem particularly trained to do conversation, so it might perform less good than training on your chat dataset.

I'll try to fork alpaca-lora to make it train on conversations, if it hasn't already been done by the time I'm on it 😅

[mouchourider]

https://github.com/tloen/alpaca-lora with a few changes seem to be a good option. (Just by changing the generate_prompt to something more chat conversation oriented than instructions oriented).

https://github.com/deep-diver/Alpaca-LoRA-Serve is also a good option: he is using the alpaca-lorca model with a prompt generated to give the instruction "give an answer to the following conversation remembering the context", but it doesn't seem particularly trained to do conversation, so it might perform less good than training on your chat dataset.

I'll try to fork alpaca-lora to make it train on conversations, if it hasn't already been done by the time I'm on it sweat_smile

I just trained the 7B model using the https://github.com/tloen/alpaca-lora project. But I don't get how you use the adapter along with the dalai project.

Does someone knows how to do it?

[Juleffel]

@mouchourider this is issue #135

[tobiasgrossmann]

https://github.com/tloen/alpaca-lora with a few changes seem to be a good option. (Just by changing the generate_prompt to something more chat conversation oriented than instructions oriented).
https://github.com/deep-diver/Alpaca-LoRA-Serve is also a good option: he is using the alpaca-lorca model with a prompt generated to give the instruction "give an answer to the following conversation remembering the context", but it doesn't seem particularly trained to do conversation, so it might perform less good than training on your chat dataset.
I'll try to fork alpaca-lora to make it train on conversations, if it hasn't already been done by the time I'm on it sweat_smile

I just trained the 7B model using the https://github.com/tloen/alpaca-lora project. But I don't get how you use the adapter along with the dalai project.

Does someone knows how to do it?

Did you found a solution? I just got the instructions together, but no idea how to get those into an adapter or model. The Alpaca one seems to be a binary. (sorry total beginner)