# anthropics/hh-rlhf

Human preference data for "Training a Helpful and Harmless Assistant with Reinforcement Learning from Human Feedback"

Repository: https://github.com/anthropics/hh-rlhf
Canonical: https://ross.abutalabs.com/products/hh-rlhf
Homepage: https://arxiv.org/abs/2204.05862
License: MIT
License Family: permissive
Archived: true
Last push: 2025-06-17T15:43:37+00:00

## Health v2 (maintenance only)
Score: 10/100 (v2, computed 2026-09-03T02:20:16.233290+00:00)
- activity 27, release rhythm 35, longevity 100
- inputs: {"age_days": 1606, "days_push": 442, "days_rel": null, "gap_med": null, "n_releases_24m": 0}
- flags: no_releases, archived
- formula: round(0.45*activity + 0.35*rhythm + 0.20*longevity); archived -> min(score, 10)

## Adoption (not part of the score)
Stars 1856, forks 160 (observed 2026-08-28T04:05:44.899321+00:00)

## What it is
Anthropic's human preference dataset for training helpful and harmless AI assistants via RLHF, plus human-generated red teaming transcripts. The GitHub repo is deprecated in favor of the HuggingFace-hosted copy of the same data.

## Use cases
- train a reward model from human preference pairs
- fine-tune an LLM with RLHF for helpfulness and harmlessness
- research alignment and preference modeling
- study red teaming attacks on language models
- evaluate harmlessness of conversational assistants
- build a safety classifier from red teaming transcripts

## When to choose
- you need the canonical Anthropic HH preference data for RLHF research
- you want red teaming conversation transcripts with harmlessness scores
- you are reproducing alignment training or preference model baselines

## When to avoid
- you need actively maintained tooling rather than raw data
- you want a dataset without potentially offensive or harmful content
- you need multilingual preference data

## Facets
- artifact type: dataset
- maturity: maintenance
- function: machine-learning, nlp, data-science
- domain: machine-learning, large-language-models, artificial-intelligence
- platform: python, cross-platform
- tags: rlhf, preference-data, red-teaming, alignment, human-feedback, l, llm-training, natural-language-processing

## Member repositories
- anthropics/hh-rlhf (main) score 10

## Provenance
- Observed fields: from GitHub, fetched 2026-08-28T04:05:44.899321+00:00.
- Health v2: computed from the inputs above; adoption is never an input.
- Inferred fields (summary, facets, guidance): AI-extracted, prompt v1, taxonomy v1, on 2026-08-30T03:16:38.223560+00:00, confidence not recorded.
  - readme: https://github.com/anthropics/hh-rlhf (fetched 2026-08-28T04:05:44.899321+00:00, sha d3270f8b1a57)
  - homepage: https://arxiv.org/abs/2204.05862 (fetched 2026-08-29T10:55:53.668577+00:00, sha 02dfcfb10487)
  - site_page: https://info.arxiv.org/about/donate.html (fetched 2026-08-29T10:55:53.677854+00:00, sha cca9c3a11c56)
  - site_page: https://info.arxiv.org/about/ourmembers.html (fetched 2026-08-29T10:55:53.682234+00:00, sha 47cbc55ff1de)
  - site_page: https://info.arxiv.org/about (fetched 2026-08-29T10:55:53.685167+00:00, sha a1f16f915a9a)
  - site_page: https://info.arxiv.org/labs/index.html (fetched 2026-08-29T10:55:53.679805+00:00, sha b14a8d05a0ec)
- Data as of 2026-08-30T08:39:29.467469+00:00.