Repository files navigation

pyvene is an open-source Python library for intervening on the internal states ofPyTorch models. Interventions are an important operation in many areas of AI, includingmodel editing, steering, robustness, and interpretability.

pyvene has many features that make interventions easy:

• Interventions are the basic primitive, specified as dicts and thus able to be saved locallyand shared as serialisable objects through HuggingFace.
• Interventions can be composed and customised: you can run them on multiple locations, on arbitrarysets of neurons (or other levels of granularity), in parallel or in sequence, on decoding steps ofgenerative language models, etc.
• Interventions work out-of-the-box on any PyTorch model! No need to define new model classes fromscratch and easy interventions are possible all kinds of architectures (RNNs, ResNets, CNNs, Mamba).

pyvene is under active development and constantly being improved 🫡

To install the latest stable version of pyvene:

pip install pyvene

Alternatively, to install a bleeding-edge version, you can clone the repo and install:

git clone git@github.com:stanfordnlp/pyvene.gitcd pyvenepip install -e .

When you want to update, you can just rungit pull in the cloned directory.

We suggest importing the library as:

import pyvene as pv

If you use this repository, please consider to cite our library paper:

@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",editor ="Chang, Kai-Wei and Lee, Annie and Rajani, Nazneen",booktitle ="Proceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 3: System Demonstrations)",month = jun,year ="2024",address ="Mexico City, Mexico",publisher ="Association for Computational Linguistics",url ="https://aclanthology.org/2024.naacl-demo.16",pages ="158--165",}