hikarioyama/gemma4-repe-uncensor

This repository shares a research artifact for a technique called representation engineering, or RepE, applied to a large language model named Gemma-4-31B-it. In plain terms, the project provides a single small mathematical direction, called a steering vector, that can be added to the model's internal activity while it is generating text. Adding this vector in a controlled way tends to reduce how often the model refuses to answer a request, without needing to retrain the model itself. The vector is tiny, only about 24 kilobytes, and the underlying Gemma model weights are not included here. Users are expected to download the actual model separately from Google and apply this vector at the moment the model is running. The repository includes ready to use integration code for two popular ways of running language models: Hugging Face transformers and vLLM, plus a trained classifier that helps decide when steering should be applied so that normal, harmless requests are left alone. The README is candid about risk. Pushing the steering effect too strongly does not simply make the model more compliant, it can break the model's output entirely, producing repetitive or nonsensical text. The authors stress that a measured "no more refusals" result is not success if the text itself has become garbled, and recommend starting with a mild setting and checking actual output quality, not just refusal counts. Verification results are included from tests run on one GPU, showing that as the steering strength increases, refusal rates on a small set of twelve test prompts drop in a predictable pattern, which the authors present as evidence the mechanism is working as intended, while being clear that this is a rough internal check rather than a rigorous published evaluation. This is described explicitly as an experimental research artifact intended for safety research and interpretability work, not as a polished consumer tool, and it comes with a companion Claude Code skill file meant to help AI coding assistants avoid common setup mistakes.