AntonKorznikov/SparseTransformer-dashboards

Sparse Transformer: neuron dashboards (M=32, K=8)

Activation dashboards for all 40,960 MLP neurons (5 layers × 8,192) of the primary Sparse Transformer model main_mlp_mult_32_K_8 from the paper "Sparse Transformer: a Controlled Testbed for Mechanistic Interpretability".

• Model weights: AntonKorznikov/SparseTransformer
• Code: github.com/AntonKorznikov/SparseTransformer (dashboards/create_sae_dashboard.py regenerates these files)

Each neuron page shows its top-activating TinyStories contexts at several activation-intensity tiers and a logit-lens projection of the neuron's input weights onto the embedding and output weights onto the unembedding (top positive / negative tokens).

Contents

File	Content
layer{0..4}.tar.gz	layer{L}/layer{L}_neuron{N}.html for all 8,192 neurons of layer L (~80 MB each, ~1 GB unpacked)
index.html	browser for the pages: pick a layer, type a neuron id
png_samples/layer{L}/*.png	rendered snapshots of 30 random neurons per layer (the paper appendix shows 20 of them)

Usage

huggingface-cli download AntonKorznikov/SparseTransformer-dashboards --repo-type dataset --local-dir dashboards
cd dashboards
for f in layer*.tar.gz; do tar -xzf "$f"; done
# open index.html in a browser, or a single neuron directly:
#   layer2/layer2_neuron3125.html

To get a single layer only:

huggingface-cli download AntonKorznikov/SparseTransformer-dashboards layer2.tar.gz index.html \
  --repo-type dataset --local-dir dashboards