Meta Llama 3 Fine-tuning, RAG, and Prompt Engineering for Drug Discovery Event Seminar and PDF 04/25/24.
Achieving the very best outputs based on most relevant specific data typically takes additional techniques beyond cutting-edge pre-trained models such as Llama 3. Two of these techniques referred to as fine-tuning and retrieval-augmented generation (RAG) can be used together with LLMs or separately alongside an LLM, depending on the objectives.
Here, three sets of experiments are detailed, with access to fine-tuned Hugging Face models, Colab links, and GitHub repositories of new developments and the authors' original notebooks. A main finding was that increasing the number of machine learning steps decreased model loss and improved the output accuracy significantly when fine-tuning on the 'Mol-Instructions' dataset. (14) A different LLM+RAG model based on the UC Irvine 'Heart Disease' dataset provided a more accurate and concise output regarding a specific directory's contents when compared to the meta.AI chatbot response. (16)
“RAG systems have a unique appeal over traditional search engines in that they can incorporate prior knowledge to fill in the gaps and extrapolate the retrieved information.”
- Kevin Wu, et al., Stanford University, April 16, 2024. How faithful are RAG models? Quantifying the tug-of-war between RAG and LLMs' internal prior