HuggingFace releases CAI recipe for open LLMs

Constitutional AI (CAI), proposed by Anthropic, is a technique in which a model critiques and revises its own answers according to a defined set of principles (Constitution) to achieve alignment. HuggingFace has released an end-to-end recipe and tools for applying this to open source models.

The main released assets are as follows:

• llm-swarm: An inference tool built on TGI and vLLM that supports large-scale synthetic data generation on Slurm clusters.
• Dataset: A harmlessness dataset built based on principles from Anthropic and Grok.
• Models: SFT and DPO models based on Mistral-7B, trained on principles from Anthropic and Grok.
• Recipe: Source code for the CAI implementation, included in the Alignment Handbook.