matrix-sae
Sparse autoencoder checkpoints for matrix-valued recurrent states. 25 checkpoints, 0.6 GB. Code and paper: github.com/JackYoung27/matrix-sae.
| group | n | size | types | description |
|---|---|---|---|---|
encoder_swap_checkpoints |
2 | 132 MB | bilinear_flat, bilinear_tied | Five-variant encoder-swap across three seeds. |
higher_rank |
2 | 16 MB | mixed | Rank-2 and rank-4 bilinear at layers 9 and 17. |
qwen3_5-0_8b_sl1024_ns2000 |
18 | 452 MB | bilinear, flat, rank1 | Main 0.8B pack: layer 9 per-head flat, rank1, bilinear. |
review_downstream_sensitivity |
2 | 8 MB | bilinear | k and nf sweeps. |
sparsity_compare |
1 | 4 MB | mixed | TopK versus BatchTopK for flat and bilinear. |
Tags: {sae_type}_L{layer}_H{head}_nf{n_features}_k{k}_s{seed}.
from loader import load_checkpoint
model, cfg, entry, _ = load_checkpoint('qwen3_5-0_8b_sl1024_ns2000/bilinear_L9_H0_nf2048_k32_s42')
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support