ParamMute RAG training examples

The ParamMute release contains 32,580 JSON Lines records for training retrieval-augmented language models. Each record includes a retrieval-conditioned input, a plain input and an output.

Which input contains retrieval?

The file uses rag_input for the retrieval-conditioned prompt, raw_input for the plain prompt and output for the target text. The filename retains the method’s earlier name, PIP-KAG. When training and evaluating the same model, keep these adaptation examples separate from the CoFaithfulQA evaluation files.

Data files

Hugging Face dataset card · The downloads below use a fixed revision.

Related paper

ParamMute: Suppressing Knowledge-Critical FFNs for Faithful Retrieval-Augmented Generation · Paper on OpenReview

Browse research data · Download catalog (JSON-LD)