igorbarshteyn/jlens-gguf — explained in plain English
Analysis updated 2026-05-18
Visualize which concepts a GGUF language model considers at each layer while generating text.
Correct a model's wrong answer by suppressing the internal representation causing the mistake.
Experiment with steering, swapping, or ablating concepts inside a running local model.
Drop in as a replacement server so any chat app gets live interpretability steering applied.
| igorbarshteyn/jlens-gguf | humanmllm/swim | juew/skill-routing-kit | |
|---|---|---|---|
| Stars | 75 | 75 | 75 |
| Language | Python | Python | Python |
| Setup difficulty | hard | hard | easy |
| Complexity | 5/5 | 5/5 | 2/5 |
| Audience | researcher | researcher | developer |
Figures from each repo's GitHub metadata at analysis time.
Requires building a native C++ activation server, a Python environment, and fitting a lens over a corpus before use.
This project is a tool for looking inside the workings of AI language models while they run, and even changing their behavior on the fly. It is built to work specifically with GGUF format models, a common way of packaging AI models to run efficiently on regular computers through a program called llama.cpp. The idea comes from an Anthropic research paper about how language models represent concepts internally. This project brings that research idea to life as an interactive visual tool. As a model processes a piece of text, this tool can show you which words or concepts the model is considering at each internal layer, and how confident it is in different possible answers. You can then reach in and adjust things while the model is thinking, for example telling it to ignore a certain concept, or nudging it toward a different one, and watch how that changes its final answer in real time. A demonstration in the readme shows the model getting confused between two similar sounding facts, giving a wrong answer. Using this tool, the author is able to identify exactly which internal representations caused the confusion and suppress just the wrong one, correcting the model's answer without retraining it. The project has two main pieces working together: a small server written in C++ that plugs directly into llama.cpp to capture and adjust the model's internal activity, and a Python and browser based interface for visualizing what is happening and applying changes. It works on models of many sizes, including large and unusually structured ones, and can even be used as a drop in replacement for a standard llama.cpp server, letting you apply these adjustments to any other application that talks to it, such as chat interfaces, while watching the effects in the visualizer at the same time.
An interactive tool to visualize and steer what's happening inside a GGUF language model as it generates text, in real time.
Mainly Python. The stack also includes Python, C++, llama.cpp.
No license information is stated in the provided README.
Setup difficulty is rated hard, with roughly 1h+ to a first successful run.
Mainly researcher.
This repo across BitVibe Labs
Verify against the repo before relying on details.