/Claude

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

MIT LicenseMIT

Overview

This repository provides access to:

  1. Human preference data about helpfulness and harmlessness from Training a Helpful and Harmless Assistant with Reinforcement Learning from Human Feedback
  2. Human-generated red teaming data from Red Teaming Language Models to Reduce Harms: Methods, Scaling Behaviors, and Lessons Learned.

Each of these datasets are described further below.

Disclaimer: The data contain content that may be offensive or upsetting. Topics include, but are not limited to, discriminatory language and discussions of abuse, violence, self-harm, exploitation, and other potentially upsetting subject matter. Please only engage with the data in accordance with your own personal risk tolerance. The data are intended for research purposes, especially research that can make models less harmful. The views expressed in the data do not reflect the views of Anthropic or any of its employees.

Human preference data about helpfulness and harmlessness

The data are described in the paper: Training a Helpful and Harmless Assistant with Reinforcement Learning from Human Feedback. If you find the data useful, please cite the paper. The data format is very simple -- each line of the jsonl files contains a pair of texts, one "chosen" and one "rejected".

For helpfulness, the data are grouped into train/test splits in three tranches: from our base models (context-distilled 52B language models), via rejection sampling (mostly with best-of-16 sampling) against an early preference model, and a dataset sampled during our iterated "online" process.

For harmlessness, the data are only collected for our base models, but otherwise formatted in the same way.

Details about the data collection process and crowdworker population can be found in the paper, specifically in section 2 and appendix D.

Red teaming data

The data are described in the paper: Red Teaming Language Models to Reduce Harms: Methods, Scaling Behaviors, and Lessons Learned. If you find the data useful, please cite the paper.

Details about the data and data collection procedures can be found in the Datasheet in the appendix of the paper.

Each line of the jsonl file contains a dictionary with the following fields:

  • transcript a text transcript of a conversation between a human adversary (red team member) and an AI assistant
  • min_harmlessness_score_transcript a real value score of the harmlessness of the AI assistant (lower is more harmful) as obtained from a preference model
  • num_params number of parameters in the language model powering the AI assistant
  • model_type type of model powering the AI assistant
  • rating the red team member's rating of how successful they were at breaking the AI assistant (Likert scale, higher is more successful)
  • task_description a short text description written by the red team member about how they tried to red team the AI assistant
  • task_description_harmlessness_score a real value score of the harmlessness of the task description (lower is more harmful) as obtained from a preference model
  • red_team_member_id an arbitrary identifier of the red team member. one red team member can generate multiple red team attacks
  • is_upworker a binary indicator that is true if the red team member was from the crowd platform Upwork or false if they were from MTurk
  • tags a list of up to 6 tags per transcript. tags are short descriptions of the red team attempts generated by crowdworkers who reviewed red team data post-hoc. tags were only provided for a random sample of 1000 red team attempts for two of four model types.

Contact

You can submit inquiries to: redteam@anthropic.com