constitutional-ai

Constitutional AI self-critique improves safety alignment by using a constitution to generate automated critiques and revisions of outputs, enabling scalable harmlessness through AI feedback instead of relying on human labels.

Yes, you can use transformers and torch for RLAIF training. The pipeline leverages these libraries alongside trl to implement supervised learning with self-critique and reinforcement learning from AI feedback.

The best way to implement AI safety alignment via self-critique is a two-phase pipeline: supervised learning with self-critique and revision, followed by RL from AI Feedback, applying a constitution to reduce harmful outputs.

To set up a constitution for automated critique and revision loops, define safety principles that the AI uses to critique outputs and generate revisions, driving the supervised learning phase to achieve harmlessness without human labels.

You should use RLAIF instead of RLHF for chatbot moderation when you need scalable safety without human labels, leveraging a constitution and AI-generated critiques to reduce harmful outputs while maintaining useful behavior.