Critique and Revision — supplies the finetuning data for → SL-CAI (Supervised-Learning Constitutional AI model)

sl-cai is not trained on some separately curated dataset; it is a pretrained model finetuned directly on the output of critique-and-revision. Concretely, the pipeline was run on 182,831 red-team prompts, sampling four critique-revision pairs per prompt from helpful-rlhf-model, and every one of those revisions (from all revision steps, not just the final one) went into the finetuning mix alongside separately collected helpfulness samples, to keep the resulting model from losing helpfulness (constitutional-ai, §"3.2 Datasets and Training", p. 8). Training ran for one epoch at a constant learning rate. This is why SL-CAI variants are named by revision depth (SL-CAI-n): the model’s behavior is a direct, traceable function of how many rounds of critique-and-revision output were included in its training data.