Visualizer•Stable
Attention-Head Visualizer
Interactive real-time visualization of multi-head self-attention weights and query-key matrix dot products.
Limitation: Simulates 8-head transformer layers with 128-dim embeddings; full 4096-dim model heads run on GPU.
Model Controls
Attention Head:Head 1 of 8
Softmax Temp ($\tau$):0.7
CURRENT LAYERLayer 12 (Self-Attn)
HEAD INDEXHead 1
AVG ENTROPY1.42 nats
MATRIX SIZE7 × 7