// the find
IntelLabs/RAG-FiT
Framework for enhancing LLMs for RAG tasks using fine-tuning.
RAG-FiT is a pipeline for fine-tuning LLMs specifically to be better RAG consumers — it handles the full loop from dataset construction (retrieval, prompt templating, few-shot assembly) through PEFT training to evaluation with RAG-specific metrics. It comes out of Intel Labs and backs a published paper. The target user is an ML engineer who wants to go beyond prompt engineering and actually train a model to ground its answers in retrieved context.
The four-stage pipeline (process → train → infer → evaluate) maps cleanly onto the actual workflow and each stage is independently runnable, so you can slot it into an existing training setup without buying into the whole thing. Hydra configuration throughout means you can sweep hyperparameters or retrieval strategies from the CLI without touching code. The evaluation module goes beyond ROUGE and EM — it integrates RAGAS and DeepEval and can run metrics over retrieval fields like citations, not just the final answer. PEFT support via TRL means you can fine-tune a 7B model on consumer hardware rather than needing a cluster.
The repo is tied to the ASQA and PubMed QA benchmarks used in the paper; adapting it to a custom domain means digging through Hydra configs and figuring out which processing steps make assumptions about those dataset schemas. The retrieval layer only ships with a Haystack/Qdrant integration — if you use any other vector store you're writing your own retriever step. Activity has been light: 61 forks and sporadic commits suggest this is closer to a research artifact than a maintained library, so don't expect upstream fixes if something breaks with newer versions of Transformers or PEFT. Documentation exists but is reference-style; there's one end-to-end tutorial (PubMed) and after that you're reading YAML files to understand what each step actually does.