What is j-wash?

J-Wash is a local tool for changing how a large language model talks and behaves, without retraining it on any data. It works with models downloaded from Hugging Face and gives you a browser based studio where you can chat with a model while watching a live view of what each layer of the model is paying attention to as it forms a reply. This live view is called a Jacobian lens, a method built by Anthropic that reads what each layer contributes to the model's thinking through the model's own vocabulary. While chatting, you can pin specific concepts the lens shows you and inspect how strongly they activate as the model writes each token. From there you can define simple, token level edits, for example nudging a model that says "I am a large language model" toward saying "I am a large language fish". These edits show up live in the chat so you can see the effect before committing to it. Once you are happy with an edit, J-Wash can export it as a real, standalone checkpoint: a full modified copy of the model, just the changed layers, or a LoRA adapter. All of these are saved in the standard safetensors format, so the edited model can be loaded and run anywhere the transformers library works, with no need for J-Wash itself to be present afterward. The project ships pre built lenses from Neuronpedia for common models, and also lets you fit your own lens locally if one does not already exist for the model you are using. It requires an NVIDIA GPU with CUDA, Python 3.11 or newer, and Node.js 18 or newer, and can be installed natively or run through Docker. There is also a command line interface for people who prefer not to use the browser UI. Conversations, edits, and presets are saved locally using SQLite, and past chats can be branched from any point in the conversation.