Gemma Scope A set of interpretability tools built to help researchers understand the inner workings of Gemma 2 Read blog Download Gemma Scope Download from Hugging Face Examine the behavior of individual model layers, while the model processes requests — to help address critical concerns including hallucinations, biases, and manipulation. View technical report Read docs Watch Overview Gemma Scope provides researchers with a suite of sparse autoencoders. Think of these as microscopes that let you zoom in on dense, compressed activations, and expand them to larger, sparser, more interpretable forms. Capabilities Perform mechanistic interpretability researchEvaluate the precise behavior of Gemma 2 models with layer-level analysis. Debug model behaviorPinpoint the source of specific model issues (such as biases and hallucinations) by examining layer-specific representations. Explore Gemma Scope Watch
