{Anthropic}

Abstract

This report includes the model card [1] for Claude models, focusing on Claude 2, along with the results of a range of safety, alignment, and capabilities evaluations. We have been iterating on the training and evaluation of Claude-type models since our first work on Reinforcement Learning from Human Feedback (RLHF) [2]; the newest Claude 2 model represents a continuous evolution from those early and less capable ‘helpful and harmless’ language assistants.This report is not intended to be a scientific paper since most aspects of training and evaluating these models have been documented in our research papers. These include papers on preference modeling [3], reinforcement learning from human feedback for helpful and harmless models [2], red teaming language models [4], measuring representation of subjective global values in language models [5], honesty, (i.e., exploring language models’ ability to recognize what they know) [6], evaluating language models with language model-generated tests [7], moral self-correction [8], and Constitutional AI [9]. We also discussed Claude’s specific constitution in a recent blog post [10]. Our work using human evaluations to test model safety is most thoroughly documented in our paper “Red-Teaming Language Models to Reduce Harms” [4], while our recent work on automated safety evaluation is “Discovering Language Model Behaviors with Model-Written Evaluations” [7]. This report is also not comprehensive – we expect to release new findings as we continue our research and evaluations of frontier models. However, we hope it provides useful insight into Claude 2’s capabilities and limitations.