For this model, we found that activations at the last token position (assuming this prompt template, with two new lines appended to the end) at layer 12 worked well.
For this model, we found that activations at the last token position (assuming this prompt template, with two new lines appended to the end) at layer 12 worked well.