Skip to main content

Hacktoberfest Hacks @ All Things Open

Live demo

MindShift

MindShift is an interactive AI interpretability lab that demonstrates how activation steering can change a language model’s behavior without fine-tuning it.

The project creates a steering direction from matched pirate-style and plain-language examples. During generation, a PyTorch forward hook adds that direction to the residual stream at a selected decoder layer.

Users can:

  • Enter any prompt
  • Select the decoder layer
  • Adjust steering strength
  • Compare baseline and steered responses side by side
  • Evaluate fluency, repetition, latency, and deterministic behavior

MindShift uses Qwen with local inference through Hugging Face Transformers. A FastAPI backend handles generation and vector loading, while a Streamlit interface provides an interactive demonstration.

The project also includes vector metadata, layer sweeps, deterministic evaluation tools, smoke tests, and automated API tests.

Built with

huggingfacepythonpytorchTransformersQwen3-1.7BfastapistreamlitpydanticpytestCUDAollamallama

AI tools

Coded withCursorOther QwenQwen3.8 Flash

Prize Categories Entered

  • DigitalOcean

    Best Open-Source AI Project

    Submitted

    MindShift uses the open-weight Qwen3-1.7B model as the core of an activation-steering lab. Using PyTorch forward hooks, we inject learned mean-difference vectors into the model’s residual stream at selected decoder layers, changing behavior without fine-tuning or prompt modification. The project includes vector construction, layer and strength tuning, deterministic evaluation, a FastAPI backend, and an interactive Streamlit comparison interface.