This repository explores the internal stochastic nature of Gemma models. By extracting transition scores and logits from the Hugging Face transformers generation loop, we can analyze the model's confidence levels and visualize "competing" tokens at each step of the sequence.