Semantics-Adaptive Dynamic Intervention (SADI)

This repository contains the data and codes for our paper "Semantics-Adaptive Activation Intervention for LLMs via Dynamic Steering Vectors".

1. Data

Please download the data of downstream tasks and put it in ./data/

2. Intervention

For Discriminative and Generative tasks:

python intervention.py

Citation

If you find this work is useful or use the data in your work, please consider cite our paper:

@article{wang2024semantics,
  title={Semantics-Adaptive Activation Intervention for LLMs via Dynamic Steering Vectors},
  author={Wang, Weixuan and Yang, Jingyuan and Peng, Wei},
  journal={arXiv preprint arXiv:2410.12299},
  year={2024}
}