LL·Mirror is a C++ GUI tool to inspect and visualize language-model activations and logits (token ranking) during generation. The app shows per-token per-layer logits, top-N token lists, heatmaps and rank plots, and provides a simple chat pane to interact with the model.
This work is inspired by Jacobian Lens and depends on lens feature of chatllm.cpp.
Highlights:
- 100% Native: no Python/JS/Web.
- Interactive visualization: fast and beautiful.
Requirements
- Windows 10/11 (x64)
Screenshots
- Download a release.
- Download any model supported by chatllm.cpp.
- Start the main program with Identity lens.
To use Jacobian Lens:
- Download a supported quantized model, for example Qwen3-4B.
- Download the pre-fitted quantized Jacobian Lens, for example Qwen3-4B.
- Start the main program with Linear lens.
-
Open the solution llmirror.sln in Visual Studio and build.
-
Copy all binding DLL of chatllm.cpp to build output directory then run.
This project uses these awesome libraries/frameworks:
This project accepts contributions (such as bug fixes), not code consignments (such as features I don't care).