Fine tuning Dolly for QA using on prem GPU
Hi, First of all thanks for doing this great job.
I have a use case where user can ask question from the given pdf file. I tested dolly its answer is decent but i need precise answer for that. So for that we need to finetune dolly. I have gone through the github repo i found codes for that but that codes are written of DB notebooks. I am new to this fine tuning thing. Please suggest how to finetune dolly on our dataset using our on prem GPU. I am having V100-SXM2-32GB gpu. I want to fine tune the 3b parameter model
Thanks for your help
You run the same code. You just need to edit out the reference to dbutils and set the choices of model, etc directly in the python code. And probably run the pip and apt installs separately, not in the python file of course. The rest is the same.
Your first step is to get text out of the PDFs and into instruction-response strings. That's up to you.
However I don't think that is going to help if your use case is answering questions over chunks of text in PDFs, because they don't contain questions.
You should try langchain for question answering on top of PDFs and an LLM (like Dolly or others). If the results aren't ideal, try tuning the generation settings.
It's hard to summarize all the details here; have a look at those things first.
Hi Thanks for quick response. I am in the process of creating training dataset as per your dataset .jasonl file, where it will have instructions along with context and answers, same way like Like Dolly 15k dataset. I hope that is correct approach. I installed all the requirements. Will try to modify the code as per ur instructions and will share code snaps if I face any issue.
Also if worked will share my findings.
Thank you
That's fine yes, if you have the same input format, everything else would work. See the notes in the README about adjusting training for V100s.
Ok sure
Where is this finetuning notebook located? I'm also interested in finetuning this model
It’s on the GitHub repo of Dolly named as train_dolly
Ah thank you!