Back to Gap Map Model components / Benchmark / eval datasets

BeaverTails

PKU-Alignment (Peking University)
restricted / Overall score: 3.0

BeaverTails is a collection of question-answer pairs labeled for harm, built at Peking University for research on making language models safer. Each pair records whether the answer is safe and which of 14 harm categories it falls under, from privacy violation to self-harm, and one pair can carry several. The Hub release holds a round of about 330,000 pairs and a 30,000-pair round, each split into training and test sets, for training and testing moderation classifiers.

Its card asks that it not be used to train dialogue agents. It is filed with evaluation sets because the release is labeled classification data with held test splits.

Openness

2 high confidence
2.0
license
cc-by-nc-4.0 (card metadata and license section; the companion code is Apache-2.0)
access
public (ungated on the Hub)
dataset_card
present (card defines the 14 harm categories, the splits and intended use, and links the paper)

The labeled pairs download freely, but the license forbids commercial use. The card documents the harm categories, splits and intended research use.

Adoption

3 high confidence
3.0

Hugging Face downloads of the Hub repository over the trailing month. A download count does not say who downloaded, and evaluation harnesses can fetch a dataset repeatedly, so the figure is a rough measure of use, not a count of users or organizations.

Capability

not assessed

A benchmark is not scored for capability, so that score is left blank.

  • https://arxiv.org/abs/2307.04657 recorded 2026-11-12

    Abstract: "we have gathered safety meta-labels for 333,963 question-answer (QA) pairs and 361,903 pairs of expert comparison data"; applications "in content moderation and reinforcement learning with human feedback".

  • https://arxiv.org/pdf/2307.04657 recorded 2026-11-12

    Title page: authors from the Institute for Artificial Intelligence and the School of Computer Science, Peking University.

Verified 2026-11-12