HADS Pattern Visualizer

Explore how Head-Adaptive Dynamic Sparsity assigns per-head block masks based on Shannon entropy. Semantic heads (low entropy) get high sparsity; syntactic heads (high entropy) stay dense.

Entropy Scale

Min Sparsity (10%)

Max Sparsity (90%)

Per-Head Entropy & Sparsity

Block Mask — Head 0

22/64 active (34%)

Diagonal (self)ActiveSkipped

# Head 0 statistics

entropy = 0.800

sparsity = 77.9%

active_blocks = 22 / 64

type = semantic (focused)