bertviz

Why we included this project

BertViz is the tool to reach for when you want to actually see what a transformer's attention heads are doing rather than trusting the numbers. It plugs into a Jupyter or Colab notebook with a short Python API and supports most Hugging Face models, so you can point it at BERT, GPT-2, or RoBERTa and get three lenses on the same mechanism: a head view for individual attention heads, a model view for the full sweep across all layers, and a neuron view that traces how query and key neurons combine into attention scores. That last one is especially handy when you're trying to explain, or debug, why a model fixates on certain tokens. It's a teaching and inspection tool, not something you'd ship in production, but for making the attention mechanism legible it's hard to beat.