Temporal Bias Analysis in Large Language Models

DateLogicQA Dataset

Overview

The DateLogicQA dataset aims to explore how large language models (LLMs) handle dates presented in different formats and contexts. It includes 190 questions divided into four categories: commonsense, factual, conceptual, and numeric. Each question uses one of seven date formats and spans three temporal contexts: past, present, and future. This systematic variation enables in‑depth analysis of LLM performance on temporal information.

Examples

• Numeric: What date is 7 years and 9 months after year 27101446?
• Factual: Which person died in the year 23041616? A) Shah Jahan B) Miguel de Cervantes C) Princess Diana D) William Shakespeare
• Conceptual: The first iPhone was released on 29062007. How many years have passed since its release?
• Commonsense: John was born on 15‑03‑1985. He graduated from university on 01‑05‑2007. Was John older than 18 at graduation?

Date Formats

• DDMMYYYY: 23041616
• MMDDYYYY: 04231616
• DDMonYYYY: 23April1616
• DD‑MM‑YY: 23-04-16
• YYYY, Mon DD: 1616, April 23
• DD/YYYY (Julian calendar): 113/1616
• YYYY/DD (Julian calendar): 1616/113

Dataset Access

The DateLogicQA dataset is available on Hugging Face at: https://huggingface.co/datasets/gagan3012/DateLogicQA.