CLIPPER: Compression enables long-context synthetic data generation

TL;DR

CLIPPER introduces a compression-based method for generating high-quality synthetic long-context data, significantly improving narrative claim verification accuracy and grounding in large language models.

Contribution

The paper presents a novel compression-based approach for synthetic data generation that enhances claim validity and reasoning complexity in long-context tasks.

Findings

Achieved 76% accuracy on narrative claim verification, a substantial improvement.

Constructed a 19K synthetic dataset with claims, source texts, and reasoning.

Set new state-of-the-art for sub-10B models on the NoCha leaderboard.

Abstract

LLM developers are increasingly reliant on synthetic data, but generating high-quality data for complex long-context reasoning tasks remains challenging. We introduce CLIPPER, a compression-based approach for generating synthetic data tailored to narrative claim verification - a task that requires reasoning over a book to verify a given claim. Instead of generating claims directly from the raw text of the book, which results in artifact-riddled claims, CLIPPER first compresses the book into chapter outlines and book summaries and then uses these intermediate representations to generate complex claims and corresponding chain-of-thoughts. Compared to naive approaches, CLIPPER produces claims that are more valid, grounded, and complex. Using CLIPPER, we construct a dataset of 19K synthetic book claims paired with their source texts and chain-of-thought reasoning, and use it to fine-tune three open-weight models. Our best model achieves breakthrough results on narrative claim verification (from 28% to 76% accuracy on our test set) and sets a new state-of-the-art for sub-10B models on the NoCha leaderboard. Further analysis shows that our models generate more detailed and grounded chain-of-thought reasoning while also improving performance on other narrative understanding tasks (e.g., NarrativeQA).