Model components / Benchmark / eval datasets

HellaSwag

A commonsense NLI and sentence-completion benchmark introduced in 2019. It contains roughly 70,000 multiple-choice examples and is widely used to test whether models can choose the plausible continuation of a sentence in everyday contexts.

HellaSwag (Zellers et al., ACL 2019), commonsense NLI sentence-completion benchmark. 59,950 examples (39,905 train / 10,042 val / 10,003 test); 4-way multiple-choice over ActivityNet-derived contexts. Canonical commonsense benchmark reported in nearly every LLM release. Verified live June 2026 on huggingface.co/datasets/Rowan/hellaswag. Registry attributes to AI2/Allen Institute; primary authors are UW/AI2 (Rowan Zellers).

Openness

5 high confidence

5.0

data: open(59,950 examples, downloadable on HF)
license: MIT(permissive, redistributable)
datasheet: yes(HF dataset card + ACL 2019 paper)
access: public(train/val/test all released, not held-out)

Fully open under MIT with dataset card; canonical open-data benchmark profile.

https://huggingface.co/datasets/Rowan/hellaswag recorded 2026-06-04
MIT license, 59,950 examples, public splits, dataset card, ~270K downloads last month

Adoption

4 high confidence

4.0

~270K HF downloads last month (>1M cumulative) and a near-ubiquitous commonsense benchmark cited across LLM release reports and the Open LLM Leaderboard. Strong usage_volume signal as an entrenched standard.

https://huggingface.co/datasets/Rowan/hellaswag recorded 2026-06-04
~270,262 downloads in the last month

Capability

not assessed

Capability not a meaningful axis for an eval dataset.

Unchanged since 2026-06-09 (last edited, not re-checked)