J-Wash
A local FastAPI and React studio for editing a language model's identity and behavior by manipulating token directions revealed by a Jacobian lens, then exporting the edits into a standalone checkpoint that runs anywhere transformers models do.
J-Wash lets a user chat with a loaded Hugging Face language model while a live Jacobian lens shows what each layer is reading from the residual stream, pin and inspect specific concepts, and then edit the model's behavior with token-level rules applied live in the chat. Once an edit looks right, it can be exported as a full checkpoint, a set of modified layers, or a LoRA adapter, producing standard safetensors weights that load with the ordinary transformers library and no J-Wash code in the loop.
It builds on Anthropic's published Jacobian lens method, which reads each layer's contribution to the residual stream through the model's own unembedding, and can use pre-fitted community lenses or lenses fitted locally. It also ships a command-line interface for the same editing pipeline without the UI.
It requires an NVIDIA GPU, Python 3.11 or newer, and Node.js, and runs natively or through Docker.