Fictitious Synthetic Data Can Improve LLM Factuality via Prerequisite Learning

TL;DR

Experiments show that Prereq-Tune outperforms existing baselines in improving LLM's factuality across short QA and long-form generation tasks and opens new possibilities for knowledge-controlled generation in LLMs.

Abstract

Recent studies have identified one aggravating factor of LLM hallucinations as the knowledge inconsistency between pre-training and fine-tuning, where unfamiliar fine-tuning data mislead the LLM to fabricate plausible but wrong outputs. In this paper, we propose a novel fine-tuning strategy called Prereq-Tune to address this knowledge inconsistency and reduce hallucinations. Fundamentally, Prereq-Tune disentangles the learning of skills and knowledge, so the model learns only the task skills without being impacted by the knowledge inconsistency. To achieve this, Prereq-Tune introduces an additional prerequisite learning stage to learn the necessary knowledge for SFT, allowing subsequent SFT to focus only on task skills. Prereq-Tune can also be combined with fictitious synthetic data to enhance the grounding of LLM outputs to their internal knowledge. Experiments show that Prereq-Tune outperforms existing baselines in improving LLM's factuality across short QA and long-form generation tasks. It also opens new possibilities for knowledge-controlled generation in LLMs. Our code is available at https://github.com/UCSB-NLP-Chang/Prereq_tune.git.

Figures (12)

• Figure 1: Overview of the proposed Prereq-Tune strategy.
• Figure 2: Procedure of creating multi-version dataset pairs for the biography generation task.
• Figure 3: Accuracy on biography generation under different numbers of generated claims.
• Figure 4: Performance as the number of synthetic data scales up.
• Figure 5: The distribution of each response type with respect to the log of monthly page views of the entities.