JsonTuning: Towards Generalizable, Robust, and Controllable Instruction Tuning

2023-10-04

arXiv.org

AI Summary

Traditional text-to-text fine-tuning suffers from inherent limitations in generalization, robustness, and output controllability. To address these issues, we propose JsonTuning—a novel structure-to-structure instruction fine-tuning paradigm that explicitly models task structure via standardized JSON. Its core innovation lies in uniformly representing input-output relationships in JSON format, thereby enhancing parseability of task elements, reducing semantic ambiguity, and enabling fine-grained control over both output format and content. We conduct systematic evaluations across multiple foundation models (e.g., LLaMA, Qwen) and diverse benchmarks—including MMLU, BIG-Bench Hard, and dedicated robustness test suites. Results show that JsonTuning achieves an average accuracy gain of +3.2% over conventional text-based fine-tuning (TextTuning), improves adversarial robustness by 17.5%, and attains a 98.4% output format compliance rate. These findings demonstrate substantial improvements in the universality, stability, and controllability of large language model fine-tuning.

Abstract

Instruction tuning is vital for enhancing the performance of large language models (LLMs), but existing text-to-text methods, referred to as TextTuning, struggle with issues such as generalization, robustness, and controllability due to their lack of explicit task structures. We introduce JsonTuning, a structure-to-structure approach that uses JSON structures to represent tasks. This method improves generalization by clarifying task elements and their relations, boosts robustness by minimizing ambiguity, and enhances controllability by allowing precise control over outputs. We conduct an extensive comparative analysis between JsonTuning and TextTuning using various language models and benchmarks. Our findings reveal that JsonTuning consistently surpasses TextTuning in terms of performance, robustness, and controllability across different scenarios. By overcoming the limitations of TextTuning, JsonTuning demonstrates significant potential for developing more effective and reliable LLMs capable of handling diverse scenarios.

Problem

• Adaptability
• Stability
• Output Control

Innovation

• JsonTuning
• Adaptability Enhancement
• Output Control