HellaSwag is a benchmark that asks a model to pick the most sensible ending to a short everyday scene, testing commonsense reasoning.
In plain words
Each question describes an ordinary situation and offers several possible next sentences. Only one makes sense to a person. The wrong endings were machine-written and then filtered so they would fool models while looking silly to humans. At release, people scored over 95% and the best models under 48%.
A quick example
Take the paper's example scene: a woman sits at a piano. Picture choices such as her placing her fingers on the keys, plus a few odd endings that sound fluent but make no sense. A person picks the right one instantly; the benchmark checks whether the model does.
Where you'll meet it
- Testing commonsense reasoning in language models.
- Research on benchmarks built with adversarial filtering.
Related
Sources
- PaperHellaSwag: Can a Machine Really Finish Your Sentence?, Zellers et al. (arXiv) · read 29 Sept 2026
- DocsHellaSwag, Rowan Zellers · read 29 Sept 2026
Checked by WhatIsAI against the sources above · 29 Sept 2026. Spotted an error? Report it.