Features
- Serialisable interventions — specified as dicts, saveable locally, and shareable via HuggingFace.
- Composable & customisable — run on multiple locations, arbitrary neuron sets, in parallel or sequence, on generative decoding steps.
- Works on any PyTorch model — no new model classes needed. Supports RNNs, ResNets, CNNs, Mamba, Transformers, and more.
Citation
@inproceedings{wu-etal-2024-pyvene,
title = "pyvene: A Library for Understanding and Improving {P}y{T}orch Models via Interventions",
author = "Wu, Zhengxuan and Geiger, Atticus and Arora, Aryaman and Huang, Jing and Wang, Zheng and Goodman, Noah and Manning, Christopher and Potts, Christopher",
booktitle = "Proceedings of NAACL-HLT 2024 (System Demonstrations)",
month = jun,
year = "2024",
publisher = "Association for Computational Linguistics",
url = "https://aclanthology.org/2024.naacl-demo.16",
pages = "158--165",
}