There's no better way to end my internship at 🐢 Giskard than by publishing an article!
In this article, I investigate the optimal number of Q&A required for fine-tuning an LLM in order to enhance a RAG system. I also study the impact of fine-tuning with distracting context on model performance.
The goal is to identify what constitues an efficient and cost-effective strategy for fine-tuning in that context.
Link: https://lnkd.in/d3WUmSrc
Key Findings:
- Adding even a single distracting document to the context when training an LLM helps it distinguish distracting context from supporting facts during inference, thus making it more efficient. This is especially interesting in a RAG setting, where the retrieval step might bring out distracting context;
- Fine-tuning Mistral-7B-Instruct for a single epoch on under 1000 training samples is very quick and can significantly improves results.
I would like to thank the entire team and especially Rabah Khalek, Benoît Malézieux, Matteo Dora, Pierre Le Jeune and Luca Martial for their guidance during those past 6 months. I learnt so much alongside the team during the various projects I participated in, especially about multimodal models, RAGs and fine-tuning.
I am extremely grateful to you guys!
So exciting! And thank you, Jayden Ramirez