Dolittle

Dolittle is a C++20-based tool paired with a browser interface that interprets and lets users interact with the internal activations of the Qwen3 8B language model via a natural language autoencoder. It supports model generation, interpretation, and editing to compare responses and understand model reasoning. The software runs on Linux with NVIDIA GPUs and macOS with Apple Silicon, utilizing CUDA and Metal, and is open-source on GitHub under Apache-2.0.