Steered Llama-v2-7b towards Effective Arguments for Liberal Readers

This is the steered Llama-v2-7b-chat-hf model.

We used the processed debateorg dataset to create the steering vectors:

We first extracted the hidden layers of effective arguments and ineffective arguments.
For each layer, from 18-20,
1. we calculate the median of the hidden vectors.
2. We substract the median of effective arguments from the median of ineffective arguments
3. We add the result to each corresponding activation layer