Transformer Self-Attention Visualizer

QKV Matrix Engine
Token Arcs & Attention Weights
Query (Q)
Key (K)
Value (V)
Softmax Attention Matrix A = Softmax(QKᵀ / √dₖ)
SEQUENCE:
QKV Vector Breakdown & Computation
Selected Query Token: Caught
Query (q) [0.42, -0.18, 0.85, 0.31]
Raw Dot Product Scores (q · kᵢ / √dₖ)
Softmax Attention Distribution (Σ = 1.0)
Target Token Raw Score Masked Exp (S / τ) Attention Weight (α)
Output Context Vector (z = Σ αᵢ vᵢ)
Context (z) [0.12, 0.94, -0.45, 0.62]

Attention State Summary

Max Attended: Caught (1.000)
Entropy: 0.00 bits
Causal Mask: Active
Enjoy this tool? Build your own with Super