Token-by-token generation — tiny-shakespeare

Sampling Tree

Every branch is a character that received nonzero probability from the model's output distribution at that step, sized and colored by that probability. The copper path is the one the sampler actually drew — drag, scroll to zoom, hover any node for its exact probability.

sampled — was also the top-probability character sampled — fell outside the shown top-k despite lower probability nonzero probability, not sampled

Built with vis-network from a real trace: at every step the model's full 65-character probability distribution was computed, its top candidates recorded, and the next character drawn by weighted random sampling — not scripted.