Downloads · 30 days
0
jakelipner/layer20-21-seed-robustness
layer20-21-seed-robustness is a machine learning model from jakelipner. Use it for the machine learning task on the model card, and read the license before you ship it in a product.
6 JumpReLU SAEs: layers {20, 21} x seeds {0, 1, 2}, all trained on instructbase with identical hyperparameters (l0=1e-3, lr=1e-3, bw=0.1, x16 expansion, 5000 steps) - only the random seed varies.
Downloads · 30 days
0
Access
Public
Updated Aug 13, 2026
Repo size
617 MB
Likes
0
Public
Click a slice to open those files.
.pt617 MB · 100%
From the Hugging Face model README
6 JumpReLU SAEs: layers {20, 21} x seeds {0, 1, 2}, all trained on instruct_base with identical hyperparameters (l0=1e-3, lr=1e-3, bw=0.1, x16 expansion, 5000 steps) - only the random seed varies.
Layer 20 expl_var: [0.9710805416107178, 0.9706024527549744, 0.9710683822631836] Layer 21 expl_var: [0.781128466129303, 0.7829174399375916, 0.7892084121704102]
Zero overlap across seeds - confirms the layer 20 vs 21 quality gap is robust.