Question about layer_id

[sorcererrandy]

Hi! When I run your notebook, I wonder how can define layer_id for emotion control?

Looking forward to hear your answers, thank you!

[justinphan3110]

Hi @sorcererrandy , layer_id and coeff are tuned parameters. In emotion_concept.ipynb and emotion_function.ipynb. We choose layer_id = list(range(-11, -30, -1)) for 13B Llama2-chat and list(range(-5, -18, -1)) for 7B Mistral/Llama2-chat

[sorcererrandy]

Thanks for your answer. I have a further question: how are the layer_id and coeff adjusted? Should I adjust the layer_id based on the accuracy after training the rep_reader, or should I adjust it based on the text output from the model?

[andyzoujm]

We usually use the set of layers that correspond to high reading accuracy, but feel free to also adjust it and coeff based on the output and your usecase.

[sorcererrandy]

I understand. Thanks!

[shiqichen17]

Hello! Could you please clarify the meaning of the "-" symbol in this context? I'm asking because I noticed its use in the code snippet from honesty_control_TQA.ipynb, there is:layer_ids = np.arange(8, 32, 3). Could you explain why there is a difference here?

[zhoushang2003]

Could @justinphan3110 kindly explain how to define the layer_id for 70B Llama2-chat in emotion_concept and emotion_function?

[shiqichen17]

I've received your e-mail , and I'll reply soon:)Shiqi Chen