Llama 2 Chat — is the rlhf tuned counterpart of → Llama 2

Llama 2 and Llama 2 Chat share an architecture and a pretraining run; what separates them is the RLHF stage applied on top to produce a model trained "using RLHF for safety and to follow human instructions in a chat format" (contrastive-activation-addition, §"3 Method", p. 3). CAA leans on that shared ancestry as an experimental design, not just background: because base and chat differ by RLHF alone and nothing else -- same weight initialization, same pretraining data, same tokenizer -- any difference found between steering vectors extracted from the two is attributable to RLHF specifically, rather than to some confound of comparing unrelated models. That is what makes base-chat-representation-comparison possible as a clean measurement rather than a confounded one. The paper generates steering vectors from the base model "to test similarity and transfer" against the chat model (contrastive-activation-addition, §"3 Method", p. 3) precisely because the pair functions as a matched before-and-after snapshot of one alignment intervention, letting internal representations be read as a record of what RLHF did, and did not, rewrite.