LLM Attention Explorer

Visualize attention matrices (layer/head) for GPT-style models. Load sample data or upload your own JSON.

Layer

Head

Tokens

Average heads Mask self

Normalize

Direction

Top-k for selected token

Upload JSON

JSON format

{
  "tokens": ["I", "love", "AI", "!"],
  "attentions": [ // layers
    [ // heads
      [[...],[...],[...],[...]], // head 0 matrix (n x n)
      [[...],[...],[...],[...]]  // head 1
    ],
    ...
  ]
}