# hellaswag by Rowan Zellers: AI Training Dataset
Source: https://savrn.com/datasets/hellaswag
Markdown alternate of the page above; the site index is https://savrn.com/llms.txt

---

## Dataset Card

HellaSwag: Can a Machine Really Finish Your Sentence? is a new dataset for commonsense NLI. A paper was published at ACL2019. An example of 'train' looks as follows. The data fields are the same among all splits. - ind: a int32 feature. - activitylabel: a string feature. - ctxa: a string feature. - ctxb: a string feature. - ctx: a string feature. - endings: a list of string features. - sourceid: a string feature. - split: a string feature. - splittype: a string feature. - label: a string feature. MIT https://github.com/rowanz/hellaswag/blob/master/LICENSE Thanks to @albertvillanova, @mariamabarham, @thomwolf, @patrickvonplaten, @lewtun for adding this dataset.

Excerpt from the card by Rowan Zellers.

## Structure

### default 59,950 rows

| Split | Rows | Size |
| --- | --- | --- |
| train | 39,905 | 41.5 MB |
| test | 10,003 | 10.4 MB |
| validation | 10,042 | 10.7 MB |

indint32activity_labelstringctx_astringctx_bstringctxstringendingslistsource_idstringsplitstringsplit_typestringlabelstring

## Details

Repository

Rowan/hellaswag

Publisher

Rowan Zellers

Task category

Not stated by the source

Tags

Not stated by the source

Size category

Not stated by the source

Languages

en

Revision

218ec52e09a7e7462a5400043bb9a69a41d06b76

Last updated

2025-07-10

## SAVRN's Notes on hellaswag

HellaSwag has been a default commonsense benchmark since its ACL 2019 paper: 59,950 four-way sentence-completion problems in 36.8 MB, split 39,905 train, 10,042 validation and 10,003 test, still pulled 444.2k times a month. The card's fields — activity labels, two-part contexts, candidate endings — tell you the shape: pick the one ending that actually continues the sentence. Five files and 36.8 MB mean storage and download are free.

Treat it as an eval, not training material: with a decade of published scores, any weight update on this test set invites benchmark-overfit suspicion. The license field on the hub record reads 'Not stated by the source', though the card text points at an MIT LICENSE file in the author's GitHub repository — pin that file if you need a clear commercial basis rather than relying on the hub metadata. License aside, train on the train split only, and say so in your model card.

SAVRN Research

## Files

5 files, 36.8 MB in total.

Data3 files · 36.8 MB

Documentation1 file · 7.0 KB

Repository1 file · 1.2 KB

Every file

| File | Type | Size | SHA-256 |
| --- | --- | --- | --- |
| data/test-00000-of-00001.parquet | Data | 6.1 MB | e572fd5579bd |
| data/train-00000-of-00001.parquet | Data | 24.4 MB | cacb12587faa |
| data/validation-00000-of-00001.parquet | Data | 6.3 MB | 899813071e1e |
| README.md | Documentation | 7.0 KB | — |
| .gitattributes | Repository | 1.2 KB | — |

## License and Download

License

Not stated by the source

Access

No access gate

[Download from Rowan Zellers](https://huggingface.co/datasets/Rowan/hellaswag)

Released by Rowan Zellers through its official repository on Hugging Face.

## Rowan Zellers

[All models and datasets](https://savrn.com/model-publishers/rowan)

## Explore More

- [The dataset directory](https://savrn.com/datasets)
- [The model directory](https://savrn.com/models)
- [Research papers](https://savrn.com/papers)

## Source

- Repository metadata, read 2026-10-19.
- [Hugging Face record](https://huggingface.co/datasets/Rowan/hellaswag)
- [How the hub is built](https://savrn.com/model-hub/methodology)