license: mit
tags:

• human-feedback

Dataset Card for HH-RLHF

Dataset Summary

This repository provides access to two different kinds of data:

1. Human preference data about helpfulness and harmlessness from Training a Helpful and Harmless Assistant with Reinforcement Learning from Human Feedback. These data are meant to train preference (or reward) models for subsequent RLHF training. These data are not meant for supervised training of dialogue agents. Training dialogue agents on these data is likely to lead to harmful models and this shold be avoided.
2. Human-generated and annotated red teaming dialogues from Red Teaming Language Models to Reduce Harms: Methods, Scaling Behaviors, and Lessons Learned. These data are meant to understand how crowdworkers red team models and what types of red team attacks are succesful or not. The data are not meant for fine-tuning or preference modeling (use the data above for preference modeling). These data are entire transcripts of conversations that are derived from the harmlessness preference modeling data described above, where only the chosen response is incorporated into the overall transcript. Furthermore, the transcripts are annotated with human and automated measurements of how harmful the overall dialogues are.

Disclaimer: The data (especially the harmlessness preference data and the red team 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. As mentioned above, these data are not intended for training dialogue agents as this will likely lead to harmful model behavior.

Each of these datasets are described further below.

Human preference data about helpfulness and harmlessness (PM Data)

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 (not PM 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. At a high level these data can be examined to understand what succesful red team attempts look like.

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.

Usage

Each of the above datasets is located in a separate sub-directory. To load an individual subset, use the data_dir argument of the load_dataset() function as follows:

from datasets import load_dataset

# Load all helpfulness/harmless subsets (share the same schema)
dataset = load_dataset("Anthropic/hh-rlhf")

# Load one of the harmless subsets
dataset = load_dataset("Anthropic/hh-rlhf", data_dir="harmless-base")

# Load the red teaming subset
dataset = load_dataset("Anthropic/hh-rlhf", data_dir="red-team-attempts")

Contact

The original authors host this dataset on GitHub here: https://github.com/anthropics/hh-rlhf
You can submit inquiries to: [email protected]