Work in progress Expect paper soon. By Arnau Marin-Llobet & Stefan Heimersheim(WIP)

Weightpedia

Weight-level interpretability for sparse transformers

191701 Weights with effect
12 Layers

View all weights (flat list) Surprise me with a random weight

Layer 0
Nonzero weights 44,777
Layer 1
Nonzero weights 44,893
Layer 2
Nonzero weights 43,860
Layer 3
Nonzero weights 44,884
Layer 4
Nonzero weights 44,921
Layer 5
Nonzero weights 43,951
Layer 6
Nonzero weights 44,973
Layer 7
Nonzero weights 45,008
Layer 8
Nonzero weights 44,006
Layer 9
Nonzero weights 45,019
Layer 10
Nonzero weights 44,051
Layer 11
Nonzero weights 45,065

Neuron Case Studies

Neuron 1863
Universal number detector (decomposition case study)
Connected weights 21
Layer 0
Neuron 863
Double-quote detector — Figure 4 string-closing circuit (Gao et al.)
Connected weights 8
Layer 0