ALTER: A Synthetic Dataset for Testing the Temporal Reasoning Capabilities of LLMs

Project Links

• Github Repository: https://github.com/MorganLFXx/Evaluation-Framework-based-Allen-Algebra

Introduction

This is a temporal reasoning test dataset automatically synthesized from an evaluation framework proposed by us, with Allen's Interval Algebra serving as the fundamental rule skeleton.

This dataset comprises three tasks, namely

• Goal-Oriented Deduction (for long-range reasoning capability)
• Local Gap Completion (for robustness under uncertainty)
• Contradiction Source Detection (for the ability to proactively identify errors)

This dataset should be used in conjunction with the code repository.

Data Statistics

This dataset contains 2500 question-answer pairs across 3 tasks, distributed by data source as follows:

• Goal-Oriented Deduction: 1000 QA pairs, including 200 pairs each for reasoning lengths of 1, 10,20,40, and 60.
• Contradiction Source Detection: 1000 QA pairs, including 200 pairs each for reasoning lengths of 1, 10,20,40, and 60.
• Local Gap Completion: 500 QA pairs, including 100 pairs each for reasoning lengths of 1, 10,20,40, and 60.

Contributions and Feedback

We welcome feedback and contributions from the research community! If you encounter any issues or have suggestions for improvements while using TIME-Lite, please:

Submit an Issue on the GitHub repository Send email to project maintainers (hkzhou@smail.nju.edu.cn)