Datasetpublic

tripathysagar/odia-hellaswag

Odia HellaSwag Odia translation of HellaSwag, a commonsense sentence-completion benchmark. Context and gold endings are translated; distractor endings in all_endings remain in English. Part of OdiaBench — parallel English–Odia benchmark translations for evaluating Odia-capable language models. Splits Split Rows test 10,003 train 39,905 validation 10,042 Total rows: 59,950 Schema Column Type Description id int64… See the full description on the dataset page: https://huggingface.co/datasets/tripathysagar/odia-hellaswag.

sourceHugging Faceapache-2.0updated 4mo agoView on Hugging Face

0likes26downloads

Dataset Card

Odia HellaSwag

Odia translation of HellaSwag, a commonsense sentence-completion benchmark. Context and gold endings are translated; distractor endings in all_endings remain in English.

Part of **OdiaBench** — parallel English–Odia benchmark translations for evaluating Odia-capable language models.

Splits

Split	Rows
`test`	10,003
`train`	39,905
`validation`	10,042

Total rows: 59,950

Schema

Column	Type	Description
`id`	int64	Pipeline row index (0-based, sorted)
`question`	string	English question / context
`answer`	string	English answer (gold ending, formatted choices, etc.)
`odia_question`	string	Odia translation of question
`odia_answer`	string	Odia translation of answer
`raw_id`	int64	Upstream example id
`activity_label`	string	HellaSwag activity category
`all_endings`	list<string>	HellaSwag four endings (English)
`label`	string	HellaSwag gold index (empty on test)

Translation methodology

OdiaBench translations were produced via Cursor subagents (Composer 2.5-fast) using the v3 translation prompt in `cursor_translate/odia_trans`:

1.Download upstream split and shard into 30-row JSONL batches
2.Translate question and answer into odia_question / odia_answer
3.Apply structural fixers, grammar audit (7 rules), and quality evaluation
4.Merge batches, dedupe by id, and export parquet splits

Translation rules preserve numbers as digits, keep benchmark-specific structural tokens (e.g. A:–D: for ARC, _ blank for Winogrande, <<calc>> / #### N for GSM8K), and transliterate proper nouns into Odia script.

Usage

python

from datasets import load_dataset

ds = load_dataset("tripathysagar/odia-hellaswag")
print(ds)
print(ds["validation"][0]["odia_question"])

Source dataset

—Original (English): Rowan/hellaswag

Citation

If you use this dataset, please cite OdiaBench and the upstream source:

bibtex

@misc{odiabench2026,
  title={OdiaBench: English-Odia Benchmark Translations},
  author={OdiaBench Contributors},
  year={2026},
  howpublished={\url{https://github.com/tripathysagar/OdiaBench}}
}

Upstream source:

bibtex

@inproceedings{zellers2019hellaswag,
  title={HellaSwag: Can a Machine Really Finish Your Sentence?},
  author={Rowan Zellers and Ari Holtzman and Yejin Choi and others},
  booktitle={ACL},
  year={2019}
}