JLens-GGUF adds interactive Jacobian-Lens visualization and steering to llama.cpp

A new tool called JLens-GGUF brings interactive Jacobian-Lens visualization and live steering capabilities to GGUF models running on llama.cpp. Inspired by Anthropic's research, this project fills a gap for GGUF-based inference engines by allowing users to observe and manipulate the internal state of their models.

• It features a native GGUF server based closely on llama.cpp that supports model observation and j-space swapping, abliteration, and steering.
• The tool works with both dense and Mixture-of-Experts (MoE) GGUF formats.
• While it can observe running llama-server models, it cannot steer them directly; steering requires the native server.
• Memory requirements for the lens scale with model size at approximately 1/8th of the model's footprint.

This implementation provides local users with access to advanced interpretability and control features previously limited to PyTorch or Hugging Face environments.