HH-RLHF (Anthropic Helpful/Harmless)
AnthropicAnthropic's human preference dataset for RLHF, holding about 169,000 rows of chosen-versus-rejected response pairs on helpfulness and harmlessness alongside red-teaming dialogues. It was released with Anthropic's papers on training helpful and harmless assistants, and is intended for preference modeling rather than supervised dialogue training.
Openness
5 high confidence- license
- mit
- card
- present
- ungated
- yes
- rows
- 169,352
- size
- 94.7MB
Publicly downloadable and ungated under the permissive MIT license.
- https://huggingface.co/api/datasets/Anthropic/hh-rlhf recorded 2026-09-29
Hub API record for Anthropic/hh-rlhf, whose tags include "license:mit" and "size_categories:100K<n<1M". Searched the body for a gated flag set to true: none.
- https://huggingface.co/datasets/Anthropic/hh-rlhf recorded 2026-09-29
HF dataset page for Anthropic/hh-rlhf listing "license:mit" among its tags, with the card summary "Human preference data about helpfulness and harmlessness". Searched the body for a gated-access prompt and an access-request form: none.
- https://huggingface.co/datasets/Anthropic/hh-rlhf/raw/main/README.md recorded 2026-09-29
Raw dataset card. Front matter reads "license: mit", the card is titled "Dataset Card for HH-RLHF", and it says the preference data are "meant to train preference (or reward) models for subsequent RLHF training".
Adoption
3 high confidenceAdoption is measured by Hugging Face downloads of Anthropic/hh-rlhf, the only usage channel on record for HH-RLHF.
- https://huggingface.co/api/datasets/Anthropic/hh-rlhf recorded 2026-08-13
32,927 downloads in the trailing 30 days for Anthropic/hh-rlhf
Capability
3 high confidenceAnthropic's papers report RLHF gains on NLP evaluations from training on this data, but newer open preference mixes leave it out, and neither Tulu 3's preference mixture nor Olmo 3's Dolci DPO set includes it.
- https://ar5iv.labs.arxiv.org/html/2204.05862 recorded 2026-09-29
ar5iv rendering of the Anthropic HH paper, including the authors' own caution that "none of our models appear particularly biased" under the BBQ-Lite metric.
- https://arxiv.org/abs/2204.05862 recorded 2026-09-29
arXiv abstract for the paper that introduced the data, stating "this alignment training improves performance on almost all NLP evaluations".
- https://huggingface.co/datasets/allenai/Dolci-Instruct-DPO recorded 2026-09-29
Dolci Instruct DPO card, describing a mixture of "125,000 pairs created with the preference heuristic described in Delta Learning". Searched the body for HH-RLHF and hh_rlhf: none.
- https://huggingface.co/datasets/allenai/llama-3.1-tulu-3-8b-preference-mixture recorded 2026-09-29
Tulu 3 8B preference mixture card, which says it "contains 272,898 generation pairs" and lists its generating models. Searched the body for HH-RLHF and hh_rlhf: none.
Verified 2026-09-27