elementmerc/senbonzakura

Senbonzakura is a research tool for modifying open-weight language models, the kind of AI models you can download and run yourself. Specifically, it targets the behavior where a model refuses to answer certain requests, and works to remove that refusal behavior directly from the model's internal weights rather than just prompting around it. The tool builds on earlier research showing that a model's tendency to refuse tends to concentrate in one main internal direction inside its neural network. This project's contribution is finding and removing several of these directions at once instead of just one, since a small number of stubborn refusals tend to hide in nearby directions that a single-direction approach misses. The README reports measurements on a specific model showing that cutting multiple directions reduces refusals further than cutting just one, while keeping the model's general coherence about the same as before. The process works in three steps. First, it studies the model's internal responses to a set of harmful and harmless example prompts to identify the directions associated with refusal. Second, it runs an automated search that tries different combinations of which internal layers to adjust and by how much, always checking that the model's ability to produce sensible text is not damaged in the process. Third, once a good configuration is found, it permanently rewrites the relevant weights so the change becomes a lasting part of the model rather than a temporary adjustment. It is installed as a Python package and used from the command line, supporting several common open-weight model families and architectures. The project includes a scoring tool to measure how often a modified model still refuses versus produces broken output, so results can be compared consistently across different runs and models. This is a specialized research tool aimed at people already working directly with open-weight AI models, not a general audience application.