#622 · Primary category: Education & Research

repeng

language-model machine-learning representation-engineering sae saes sparse-autoencoder sparse-autoencoders transformers

A library for making RepE control vectors

Project last updated:09/24/25

GitHub Stars

757

Forks

67

Contributors

5

License

MIT

Why we included this project

repeng is a small Python library that makes representation engineering practical. You wrap a Hugging Face transformer model, build a short dataset of paired statements that differ along a trait you care about, and train a control vector in under a minute. At inference time you apply that vector to steer generation, say making a model answer as if it were more or less 'trippy' or honest. The API stays small, a few classes and a train call, with torch and transformers doing the heavy lifting underneath, so it suits people who want to steer model behavior without building the machinery from scratch. The included notebooks walk through the whole workflow, from dataset construction to vector export.

Articles for this project

No articles for this project yet.

To suggest a topic or contribute an article, contact us.

Related projects in this category