Learning to Verify Summary Facts with Fine-Grained LLM Feedback

TL;DR

This paper presents FineSumFact, a large-scale dataset of fine-grained factual feedback on summaries generated by multiple LLMs, used to fine-tune a lightweight model that outperforms human-annotated data in fact verification tasks.

Contribution

Introduces FineSumFact, a novel dataset leveraging LLM-generated feedback for training summary fact verifiers, improving efficiency and performance over human-labeled data.

Findings

Model trained on LLM-generated data outperforms human-labeled data-based models.

Fine-tuning with LLM feedback is more cost-effective than human feedback.

The dataset enables high-performance fact verification with lightweight models.

Abstract

Training automatic summary fact verifiers often faces the challenge of a lack of human-labeled data. In this paper, we explore alternative way of leveraging Large Language Model (LLM) generated feedback to address the inherent limitation of using human-labeled data. We introduce FineSumFact, a large-scale dataset containing fine-grained factual feedback on summaries. We employ 10 distinct LLMs for diverse summary generation and Llama-3-70B-Instruct for feedback. We utilize this dataset to fine-tune the lightweight open-source model Llama-3-8B-Instruct, optimizing resource efficiency while maintaining high performance. Our experimental results reveal that the model trained on extensive LLM-generated datasets surpasses that trained on smaller human-annotated datasets when evaluated using human-generated test sets. Fine-tuning fact verification models with LLM feedback can be more effective and cost-efficient than using human feedback. The dataset is available at https://github.com/DISL-Lab/FineSumFact.