Skip to content

Dataset · Eureka-Lab

Physics perception and reasoning benchmark for LLMs

Eureka-Lab/PHYBench

English text-based tasks for evaluating physics understanding in large language models.

What's inside

Contains English text tasks to test physics perception and logical reasoning. The dataset size ranges from 1,000 to 10,000 records. The authors also released the EED Score metric code for answer quality evaluation and maintain a leaderboard.

What you can do with it

  • Evaluating LLM ability to solve physics problems and perform reasoning
  • Comparing models using the EED Score metric
  • Analyzing language model error patterns in physical process understanding tasks
Tasks
Question answering
Languages
English
Size
1K–10K rows
Downloads this month
3.3K
Likes
62

Models trained on it

How to load it

from datasets import load_dataset

ds = load_dataset("Eureka-Lab/PHYBench")