Comment by Siyu on Mechanistically Eliciting Latent Behaviors in Language Models · 2024-10-09T13:44:53.986Z · LW · GW

Hi Andrew,

Thank you for this amazing post. I have a question about the application. For each dataset, such as 'bombing' and 'chain of thought', when training the LLM's vectors, do you construct objective "examples" with a specified 'Q:' for the prompt and a targeted 'A:' for the model to learn the desired behavior? I've noticed that all the examples in the notebook only contain one prompt and answer. If so, how many data points do you have in the examples for training? thank you very much for your help and I look forward to hearing from you!