SNIP Scope
SNIP Scope trains a top-k sparse autoencoder on final-layer hidden activations from the pretrained SNIP-0.4M transformer. A 96-dimensional activation is encoded into a 384-feature overcomplete dictionary, but only the 16 largest positive features may fire for each token.
The benchmark measures held-out reconstruction, explained variance, active-feature count, dead-feature rate, and token exemplars for each learned feature. A 16-component PCA reconstruction is retained as a dense low-rank control. Feature exemplars are descriptive clues, not proof that a neuron represents one human concept.
Verified results
The SAE trained on 180,000 final-layer token activations and was measured on 40,000 held-out activations.
| Metric | Top-k SAE | PCA-16 control |
|---|---|---|
| Reconstruction MSE | 0.02585 | 0.29824 |
| Explained variance | 97.41% | 70.11% |
| Active features per token | 16.00 | 16 dense components |
| Dead dictionary features | 3.39% | not applicable |
The learned dictionary contains 384 features and the SAE has 74,112 parameters. The Space exposes each feature's five highest-activating held-out BPE tokens and firing rate. Those exemplars may reflect token identity, syntax, position, or mixed causes; the project does not assign automatic human-readable concepts.
Reproduce
uv run python projects/snip-scope/train.py