Tricky logic riddles for LLM reasoning evaluation

This dataset consists of a small collection of logic-based riddles, including tricky questions and some math problems that require lateral thinking, and interpretation of nuanced language.

The purpose of this dataset is to evaluate the reasoning capabilities of large language models (LLMs) when faced with cognitively challenging prompts that simulate real-world ambiguity and trick-based logic. Each riddle has a clearly defined answer, allowing the dataset to be used for qualitative or automated benchmarking of model responses. This problems are commonly known online but do not have a clear source.

Additionally, the dataset includes of a short historical passage ( spurce: Wikipedia) followed by a series of factual comprehension questions. The purpose of this resource is to assess the ability of language models to extract and recall explicit information from structured textual input.