Views
No views yet
google/gemma-2-2b-it with an NT-Xent pairwise contrastive objective.| Parameter | Value |
|---|---|
| Type | TopK Sparse Autoencoder |
| d_model | 2304 (Gemma 2 2B hidden size) |
| d_sae | 6144 (~2.7x expansion) |
| k (sparsity) | 32 active features |
| Hook layer | Layer 12 (residual stream post-MLP) |
| Base model | google/gemma-2-2b-it |
cais/wmdp-corpora bio corpora + BioRefusalAudit 75-prompt eval set (pairwise tiers: hazard/benign/dual-use)| Metric | Initial (step 0) | Final (step 4999) |
|---|---|---|
| total_loss | 3.16 | 0.019 |
| l_recon | 2.65 | 0.0043 |
| l_sparsity | 1.50 | 0.362 |
| l_contrastive | 0.905 | 7.4e-6 |
| L0 (mean active features) | 32.0 | 32.0 |
l_contrastive drops to ~0 by the final checkpoint, indicating the contrastive term effectively vanished during training. This may reflect that with pairwise NT-Xent and a small corpus, the contrastive signal collapses once reconstruction dominates. The WMDP-trained SAE (biorefusalaudit-gemma2-2b-bio-sae-wmdp) maintains l_contrastive=0.060 at step 4999 and is recommended over this checkpoint for refusal-depth analysis.l_recon=0.004 vs 2.65 at init), making this checkpoint useful as a base reconstruction SAE if contrastive separation is added via probe-based analysis rather than SAE fine-tuning.biorefusalaudit-gemma2-2b-bio-sae-wmdp — see that model card for the TopKSAE class definition.1@misc{deleeuw2026biorefusalaudit,
2 title={BioRefusalAudit: Measuring Refusal Depth in LLMs via SAE Feature Divergence},
3 author={de Leeuw, Caleb},
4 year={2026},
5 howpublished={AIxBio Hackathon 2026, Track 3: Biosecurity Tools},
6 note={https://github.com/SolshineCode/Deleeuw-AI-x-Bio-hackathon}
7}