# HellaSwag

> HellaSwag is a commonsense reasoning benchmark in which a model must pick the most plausible continuation of a description of an everyday situation from four options, with wrong options generated and filtered to fool models.

- Identifier: PTL-0195
- Category: Benchmarks
- Canonical URL: https://protologue.com/t/hellaswag/
- Introduced: 2019

## Description

When it was introduced, the questions were trivial for people but hard for the best models. Modern models score close to human level, so it is a standard example of a saturated benchmark, though it still appears in model reports.

## Broader terms

- [Benchmark](https://protologue.com/t/benchmark/)

## Related terms

- [Benchmark Saturation](https://protologue.com/t/benchmark-saturation/)

## Sources

- Zellers et al. (2019). HellaSwag: Can a Machine Really Finish Your Sentence?. https://arxiv.org/abs/1905.07830

## Cite this entry

Protologue. (2026). HellaSwag. In Protologue: A Taxonomy of Prompting and LLM Techniques (v1.4.0, PTL-0195). https://protologue.com/t/hellaswag/

License: CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/)