Taking machine learning apart, one notebook at a time.
Top-K keeps a fixed number of words every time. Top-P keeps fewer when the model is sure and more when it isn't.
Read notebook →The model ranks thousands of words to pick just one. Top-K keeps a few and drops the rest.
Read notebook →I turned one dial and watched a model go from boring and predictable to weird and creative. Here's what temperature actually does.
Read notebook →Taking a small language model apart and rebuilding text generation by hand, one step at a time.
Read notebook →