J-Wash is a local studio that lets you edit a Hugging Face LLM's identity and behavior by modifying token directions via a Jacobian lens, then export the result as a real checkpoint.

Stars

201

7-day growth

No data

Forks

19

Open issues

2

License

Apache-2.0

Last updated

2026-07-20

AI repository intelligence
FR-AI / ANALYSIS

Why it is worth attention

It enables no-training, no-dataset model editing with a live preview of changes, and exports standard safetensors weights that can be used anywhere without additional code.

Who it is for

  • ML researchers exploring interpretability
  • developers wanting to customize LLMs without fine-tuning
  • AI safety practitioners testing behavioral modifications
  • hobbyists interested in hands-on model editing

Use cases

  • Changing a model's persona (e.g., from assistant to fish)
  • Removing or attenuating specific unwanted behaviors
  • Creating custom personalities for characters or applications
  • Exploring how different layers contribute to model outputs

Strengths

  • Live Jacobian lens visualization with heatmaps, token clouds, and per-layer rank curves
  • Supports multiple export formats: full checkpoint, modified layers, LoRA, and GGUF
  • Edits can be previewed in real-time during chat and exported faithfully
  • Works with any Hugging Face transformers model, including fine-tunes on their base model lens

Considerations

  • Requires an NVIDIA GPU with CUDA; no CPU fallback
  • Only supports safetensors/transformers models, not GGUF files directly
  • Editing effectiveness varies by architecture (Gemma models are harder to modify)
  • May degrade model intelligence; requires testing and intuition to avoid dumbing down

README quick start

Installation

Description

Jacobian-Brainwash : A manual alignment tool for large language models built on Anthropic's Jacobian Lens. Results are exportable.

Related repositories

Similar projects matched by category, topics, and programming language.

0xwilliamortiz
Featured
0xwilliamortiz GitHub avatar

openclaude-improved

OpenClaude is an open-source CLI coding agent that runs on any platform and supports a wide range of LLM providers, offering the same tools and workflows as Claude Code.

AI & Machine LearningLarge Language Models
577
deerwork-ai
Featured
deerwork-ai GitHub avatar

deer-workflow

An open-source Dynamic Workflow runtime that combines deterministic TypeScript orchestration with replaceable Agent runtimes.

AI & Machine LearningLarge Language Models
312
S40911120
Featured
S40911120 GitHub avatar

recensa

Recensa is a self-hosted web viewer that indexes Claude Code session transcripts into a local SQLite database, enabling full-text search, replay, and audit of all past agent conversations without uploading data anywhere.

AI & Machine LearningLarge Language Models
67