RAG-FiT

Why we included this project

Retrieval-augmented systems usually pair a general LLM with a retrieval pipeline, and the model often doesn't follow the exact retrieval, ranking, and prompt conventions that pipeline uses. RAG-FiT addresses this by fine-tuning the model on datasets built to mirror your own RAG setup, so it learns your patterns instead of generic ones. The library covers the whole workflow: it generates RAG-augmented training data from the technique you choose, trains with parameter-efficient methods, and evaluates the results with RAG-specific metrics like RAGAS and BERTScore. Since the pipeline is modular and driven by configuration files, you can experiment with different retrieval and prompt setups without rewriting code. That's useful for teams that want more grounded, accurate answers from an existing model without building a full training stack from scratch.