Show HN: LLM Attention Visualization

The tool visualizes transformer attention by letting users hover over generated tokens to see which prior tokens contributed most, using attention weights scaled by value-vector magnitude aggregated across heads and layers. Demonstrations show verbatim copying of addresses, date strings, and combined word influences, even with a 600-million-parameter model, while the implementation relies on a custom React app, Transformers.js, and modified ONNX files to expose internal values.