layer int64 0 8 | seed int64 0 0 | model_id stringclasses 1
value | d_in int64 1.54k 1.54k | n_features int64 49.2k 49.2k | k int64 500 500 | batch_tokens int64 32.8k 32.8k | n_steps int64 3k 15k | lr float64 0 0 | lambda_l0 float64 0 0.01 | tier stringclasses 1
value | tier_lambda_init float64 0 0 | scheduler_mode stringclasses 3
values | scheduler_transitions int64 0 6 | scheduler_l0_adjustments int64 0 0 | resample_every int64 5k 5k | bluey_frac float64 0 0 | total_tokens int64 98.3M 492M | early_stopped bool 2
classes | final_metrics dict | training_curve dict |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
0 | 0 | google/gemma-4-E2B-it | 1,536 | 49,152 | 500 | 32,768 | 3,501 | 0.0002 | 0.001455 | hot | 0.001 | BASELINE | 0 | 0 | 5,000 | 0 | 114,720,768 | true | {
"recon_loss": 0.007516,
"mean_l0": 493.03,
"dead_pct": 0,
"resampled": 0,
"ev": 0.913284300405461,
"nonlinear_err": 0.932134,
"linear_err": 3.139587
} | {
"recon_loss": [
0.108335,
0.024691,
0.019342,
0.017523,
0.018513,
0.016779,
0.015388,
0.01328,
0.011073,
0.009697,
0.008904,
0.008301,
0.007906,
0.007516
],
"mean_l0": [
1783.5,
534.17,
408.47,
315.38,
274.76,
275.32,
298.88,
... |
1 | 0 | google/gemma-4-E2B-it | 1,536 | 49,152 | 500 | 32,768 | 3,001 | 0.0002 | 0.001254 | hot | 0.001 | BASELINE | 0 | 0 | 5,000 | 0 | 98,336,768 | true | {
"recon_loss": 0.006778,
"mean_l0": 518.47,
"dead_pct": 0,
"resampled": 0,
"ev": 0.8802684129021492,
"nonlinear_err": 0.619093,
"linear_err": 3.081543
} | {
"recon_loss": [
0.052496,
0.021468,
0.01927,
0.017214,
0.016526,
0.013918,
0.011446,
0.008943,
0.007916,
0.007316,
0.007011,
0.006778
],
"mean_l0": [
1049.95,
329.1,
298.76,
258.22,
263.14,
310.25,
413.37,
504.73,
527.84,
52... |
2 | 0 | google/gemma-4-E2B-it | 1,536 | 49,152 | 500 | 32,768 | 9,501 | 0.0002 | 0.003252 | hot | 0.001 | STABILIZE | 1 | 0 | 5,000 | 0 | 311,328,768 | true | {
"recon_loss": 0.022394,
"mean_l0": 493.44,
"dead_pct": 33.75,
"resampled": 0,
"ev": 0.9350029766564396,
"nonlinear_err": 3.888109,
"linear_err": 4.166528
} | {
"recon_loss": [
7.465517,
0.077331,
0.062948,
0.057014,
0.056844,
0.052166,
0.053455,
0.050392,
0.042427,
0.027514,
0.018477,
0.013523,
0.010874,
0.008848,
0.007865,
0.007259,
0.006537,
0.006502,
0.007011,
0.00776,
0.008835,
0.0... |
3 | 0 | google/gemma-4-E2B-it | 1,536 | 49,152 | 500 | 32,768 | 15,000 | 0.0002 | 0.005 | hot | 0.001 | RECOVERY | 6 | 0 | 5,000 | 0 | 491,520,000 | false | {
"recon_loss": 0.191992,
"mean_l0": 11374.79,
"dead_pct": 39.91,
"resampled": 19056,
"ev": 0.7510223542848521,
"nonlinear_err": 13.979935,
"linear_err": 6.642598
} | {
"recon_loss": [
10.902876,
0.108148,
0.085059,
0.075611,
0.076275,
0.070479,
0.063956,
0.042471,
0.025714,
0.016418,
0.011203,
0.00776,
0.006002,
0.004785,
0.004531,
0.004416,
0.004687,
0.005254,
0.006591,
0.008107,
0.009771,
0.... |
4 | 0 | google/gemma-4-E2B-it | 1,536 | 49,152 | 500 | 32,768 | 7,001 | 0.0002 | 0.00475 | hot | 0.001 | BASELINE | 0 | 0 | 5,000 | 0 | 229,408,768 | true | {
"recon_loss": 0.028522,
"mean_l0": 805.94,
"dead_pct": 1.13,
"resampled": 0,
"ev": 0.9654411768423038,
"nonlinear_err": 5.411888,
"linear_err": 3.605405
} | {
"recon_loss": [
30.839931,
0.136455,
0.103278,
0.088517,
0.0906,
0.082933,
0.087842,
0.089789,
0.090479,
0.085299,
0.075902,
0.056139,
0.038968,
0.025846,
0.018414,
0.013779,
0.011239,
0.009852,
0.010066,
0.010778,
0.011949,
0.0... |
5 | 0 | google/gemma-4-E2B-it | 1,536 | 49,152 | 500 | 32,768 | 7,501 | 0.0002 | 0.005 | hot | 0.001 | BASELINE | 0 | 0 | 5,000 | 0 | 245,792,768 | true | {
"recon_loss": 0.034168,
"mean_l0": 887.93,
"dead_pct": 0.01,
"resampled": 0,
"ev": 0.9655632377736252,
"nonlinear_err": 5.82934,
"linear_err": 4.029906
} | {
"recon_loss": [
44.121544,
0.21373,
0.143746,
0.11873,
0.117568,
0.103543,
0.107503,
0.105408,
0.093981,
0.065946,
0.044339,
0.028195,
0.020395,
0.016011,
0.01417,
0.012575,
0.014026,
0.011216,
0.011656,
0.012492,
0.01321,
0.016... |
6 | 0 | google/gemma-4-E2B-it | 1,536 | 49,152 | 500 | 32,768 | 8,001 | 0.0002 | 0.005 | hot | 0.001 | BASELINE | 0 | 0 | 5,000 | 0 | 262,176,768 | true | {
"recon_loss": 0.03262,
"mean_l0": 867.6,
"dead_pct": 0.11,
"resampled": 0,
"ev": 0.9636464625006718,
"nonlinear_err": 5.443166,
"linear_err": 4.265441
} | {
"recon_loss": [
44.599266,
0.302845,
0.163016,
0.129583,
0.12195,
0.105162,
0.103257,
0.092319,
0.06989,
0.044782,
0.034729,
0.026555,
0.022397,
0.019188,
0.017911,
0.016271,
0.015855,
0.013832,
0.01381,
0.013779,
0.013515,
0.01... |
7 | 0 | google/gemma-4-E2B-it | 1,536 | 49,152 | 500 | 32,768 | 9,251 | 0.0002 | 0.005 | hot | 0.001 | BASELINE | 0 | 0 | 5,000 | 0 | 303,136,768 | true | {
"recon_loss": 0.028981,
"mean_l0": 1141.12,
"dead_pct": 0.74,
"resampled": 0,
"ev": 0.9748892890172826,
"nonlinear_err": 5.433137,
"linear_err": 3.623924
} | {
"recon_loss": [
69.262444,
2.519725,
0.354593,
0.198537,
0.157914,
0.10738,
0.064818,
0.037944,
0.02751,
0.020223,
0.016993,
0.01374,
0.012306,
0.01062,
0.010767,
0.009886,
0.008799,
0.008084,
0.007969,
0.007811,
0.007815,
0.008... |
8 | 0 | google/gemma-4-E2B-it | 1,536 | 49,152 | 500 | 32,768 | 12,501 | 0.0002 | 0.005 | hot | 0.001 | STABILIZE | 1 | 0 | 5,000 | 0 | 409,632,768 | true | {
"recon_loss": 0.019313,
"mean_l0": 1151.09,
"dead_pct": 1.73,
"resampled": 1,
"ev": 0.9723314116506296,
"nonlinear_err": 3.485507,
"linear_err": 4.02997
} | {
"recon_loss": [
49.737839,
0.617068,
0.256953,
0.17001,
0.139904,
0.106723,
0.083446,
0.058557,
0.042705,
0.030123,
0.023651,
0.019158,
0.016647,
0.015243,
0.014548,
0.013261,
0.012329,
0.011488,
0.011693,
0.014419,
0.011383,
0.... |
YAML Metadata Warning:empty or missing yaml metadata in repo card
Check out the documentation for more information.
Gemma-4-E2B SAE Atlas
JumpReLU Sparse Autoencoders trained on every residual stream layer of google/gemma-4-E2B-it using an adaptive Lagrangian controller that eliminates manual per-layer hyperparameter tuning.
What this is
A complete layer-by-layer SAE atlas for Gemma-4-E2B, trained and published live as each layer completes. Each SAE decomposes the residual stream activations at that layer into a sparse dictionary of 49,152 learned features.
Prior work
Before training these SAEs, the model was mapped behaviorally using a neural census pipeline across 35 layers × 8 components × 16 behavior categories on 184,320 probe prompts. The results are interactive and fully queryable.
The atlas identified several structural findings that informed SAE training priorities, including a three-phase behavioral leadership transition in the first four layers, a deep-layer gate sparsification event at L23-26, and a selectivity plateau at L4-L6 where neurons are 6.7× more likely to be category-selective than topic-entangled.
Training methodology
Standard JumpReLU SAE architecture with an adaptive Augmented Lagrangian controller for automatic sparsity targeting. The controller treats L0 sparsity as a hard constraint rather than a soft penalty, using projected dual ascent to find each layer's natural KKT-point without manual tuning.
The key finding: every layer converges to a different λ equilibrium automatically. No grid search, no failed runs, no human in the loop.
Results so far
| Layer | EV (best) | L0 (best) | dead% (final) | λ_eq | steps | status |
|---|---|---|---|---|---|---|
| 0 | 0.913 | 493 | 0.0% | 1.455e-3 | 3,501 | ✅ |
| 1 | 0.880 | 518 | 0.0% | 1.254e-3 | 3,001 | ✅ |
| 2 | 0.981 | 847 | 3.9% | 3.260e-3 | 9,501 | ✅ use best ckpt |
| 3 | 0.994 | 1112 | — | 4.650e-3 | 15,000 | ⚠️ unstable — use best ckpt |
| 4 | 0.988 | 1111 | 1.1% | 4.750e-3 | 7,001 | ✅ use best ckpt |
| 5 | 0.989 | 1217 | 0.01% | 5.000e-3 | 7,501 | ✅ use best ckpt |
| 6 | — | — | — | — | — | 🔄 training |
| 7-9 | ⏳ queued | |||||
| 10-34 | ⏳ queued |
Layer 2 is notable — it required 2.2× more steps and a 2.2× higher λ equilibrium than layer 0, consistent with the entanglement cliff measured independently in the brain atlas. The controller handled it automatically.
Architecture
- Dictionary size: 49,152 features (24× overcomplete)
- Activation: JumpReLU with learned per-feature thresholds
- L0 target: 500 active features per forward pass
- Training data: FineWeb-Edu (pre-tokenized)
- Base model:
google/gemma-4-E2B-it
Paper
Methodology writeup coming. The short version: this approach makes full-model SAE atlas training accessible on a single A100 for under $20 total, with zero manual tuning per layer.
- Downloads last month
- 32