
by Anthropic
HH-RLHF is Anthropic's dataset of human preference comparisons on helpfulness and harmlessness, plus red teaming transcripts. It is meant for research on making models less harmful.
DatasetFree
Step 1: Read the dataset card disclaimer about offensive content before opening the data.
Step 2: Install the datasets library and call load_dataset("Anthropic/hh-rlhf") to load the helpfulness and harmlessness subsets.
Step 3: Pass data_dir="harmless-base" to load one harmless subset.
Step 4: Pass data_dir="red-team-attempts" to load the red teaming subset.
Step 5: Use the preference pairs for reward model training and the red team transcripts for analysis, and cite the matching paper.
The preference data trains preference or reward models for RLHF. The red teaming data helps researchers understand how red team attacks work.
The card says not to. Training dialogue agents on this data is likely to lead to harmful models.
Yes. The card says it includes discriminatory language and discussions of abuse, violence, self-harm and exploitation, so use it according to your own risk tolerance.
Call load_dataset with data_dir set to red-team-attempts.
Claude in ChromeClaude in Chrome is a browser extension that lets Claude read the page you're on, then click, type, navigate and fill in forms for you. It is for paid Claude users who want to hand off browser tasks.
Claude DesignClaude Design turns a written description into a first version of a visual, such as a mockup, one-pager or slide, which you then refine with Claude. It is for people who need on-brand designs without waiting on a design team.
Claude for Microsoft 365Claude for Microsoft 365 puts Claude in a sidebar inside Excel, PowerPoint, Word and Outlook, with context carried across the four apps. Edits stay reviewable before anything is saved or sent.