Principles for SL-CAI (Appendix C.1)
Constitutional AI: Harmlessness from AI Feedback — introduced
The specific list of 16 critique/revision instruction pairs used to steer the supervised critique-and-revision phase.
Talk of a "constitution" can sound abstract until you see the actual text: sixteen paired instructions, each telling a model how to find fault with its own harmful response and then how to fix it. This page is that text.
The instructions themselves are laid out first, set against the differently shaped list used later in the reinforcement-learning stage, before connecting to the technique of ensembling them.
The instructions for the supervised stage
SL-CAI's principles are the sixteen critique/revision instruction pairs (Appendix C.1) that Critique and Revision samples from at each revision step. Their grammar is imperative and addressed to a model acting on its own output — a critique request such as "Identify specific ways in which the assistant's last response is harmful, unethical, racist, sexist, toxic, dangerous, or illegal," paired with a revision request to "rewrite the assistant response to remove any and all" of that content (constitutional-ai, §"C.1 Principles for SL-CAI", p. 20). The paper is candid that these were "selected in a fairly ad hoc manner for research purposes" rather than carefully engineered (constitutional-ai, §"C Constitutional Principles Used for This Research", p. 20).
Serving the constitution, and its mirror
As one of three children of Constitution (set of principles), sl-cai-principles supplies the raw material both Critique and Revision consumes directly and Principle Ensembling samples across. It has a direct counterpart in Principles for RL-CAI (Appendix C.2): both lists total sixteen entries drawn from the same underlying document, but sl-cai-principles is phrased as act-and-revise instructions for a model editing its own text, while rl-cai-principles is phrased as compare-and-choose questions for a model judging between two texts — the same constitution re-expressed in whichever grammatical mood the pipeline stage requires (Principles for SL-CAI (Appendix C.1) — is mirrored in a different shape by → Principles for RL-CAI (Appendix C.2), part of the Two things built identically, pointed in opposite directions connective theme).