- https://arxiv.org/pdf/2403.10131
- Source: arxiv
TL;DR
RAFT (Retrieval Augmented Fine-Tuning) trains LLMs to robustly ignore distractor documents in retrieved context by exposing them to mixed oracle and distractor passages with chain-of-thought rationales during supervised fine-tuning.
Summary
Standard fine-tuning on clean documents does not prepare LLMs for the noisy context produced by real-world retrieval. RAFT closes this gap by training on mixed sets of oracle documents and distractor passages, teaching the model to identify and cite the correct source while ignoring irrelevant content. Chain-of-thought rationales in the training data make this reasoning explicit. RAFT outperforms both RAG-only and SFT-only baselines on domain-specific benchmarks PubMed, HotpotQA, and Gorilla, and is retriever-agnostic since only the generation stage is fine-tuned.
Key contributions
- Proposes RAFT, a fine-tuning recipe that exposes LLMs to mixed oracle and distractor documents to build robustness to noisy retrieval.
- Introduces chain-of-thought rationale supervision that teaches the model to explicitly reason through which document supports the answer, providing an implicit citation mechanism.
- Demonstrates that standard SFT on clean documents does not transfer to RAG settings and retrieval-context training is required.
- Shows RAFT is retriever-agnostic by fine-tuning only the generation stage while leaving the retriever unchanged.
When to cite
- When applying domain-specific fine-tuning to a RAG system over a known closed corpus.
- When arguing that models must be trained on retrieved-context inputs (not clean documents) to perform well with retrieval at inference.
- When using distractor training to improve LLM robustness to imperfect retrieval precision.
- When justifying chain-of-thought supervision as a method for improving answer attribution and interpretability in RAG.