Scientific question-answering by fine-tuned large language models (LLMs)

Create an environment using the following command:

conda env create --file env.yml

Dataset collection

• Galaxy help forum
• Biostars Q&A

Fine-tune Llama2 (2B and 7B)

• Navigate to \llama2 and then execute python qlora-train.py
• Utilizes HuggingFace's Transformers package to download pre-trained LLMs
• qLoRA to drastically reduce the number of parameters (from 2B to 6 million)
• SFT for setting up the training process

Outcomes

Retrieval augmented generation (RAG)

• Dataset collection from Galaxy's training material and GitHub pull requests
• https://github.com/uwwint/discourse-scraper/blob/master/extract_documents/create_RAG_docs.ipynb
• Pipeline for RAG that uses fine-tuned Llama2 using Haystack

Save the fine-tuned model to HuggingFace Hub

• Save model to HuggingFace Hub: https://github.com/uwwint/discourse-scraper/blob/master/llama2/save_to_hub.ipynb
• Model name: anuprulez/fine-tuned-gllm