juiceb0xc0de/minicpm5-1b-SAE

One JumpReLU SAE per layer of MiniCPM5-1B. All 24 layers, complete.

---

MiniCPM5-1B: 24 layers, 1536-dim residual stream, 130,560-token bilingual vocab.

Every SAE in this repo: d_in=1536, 49,152 features (32x expansion), JumpReLU activation, streamed FineWeb-Edu, target sparsity L0=50. Same settings on every layer, no hyperparameter changes were applied in the run.

Each layer_NN_s0/ holds:

• sae.pt - the weights
• meta.json - config and final metrics
• checkpoint_full.pt - full optimizer state

The numbers

Layer	EV	L0	Dead	Residual RMS
0	0.650	48.91	0.00%	0.133
1	0.621	48.81	0.01%	0.172
2	0.648	48.61	0.04%	0.228
3	0.728	49.16	0.43%	0.327
4	0.990	50.18	0.06%	2.301
5	0.988	47.71	0.11%	2.372
6	0.987	49.77	0.22%	2.544
7	0.983	48.82	0.07%	2.356
8	0.983	50.41	0.09%	2.382
9	0.972	49.58	0.09%	2.408
10	0.970	50.22	0.58%	2.650
11	0.963	49.55	0.09%	2.493
12	0.954	47.69	0.09%	2.581
13	0.952	49.27	0.85%	2.765
14	0.948	48.32	0.29%	2.848
15	0.933	48.13	0.30%	3.061
16	0.924	47.14	0.03%	3.413
17	0.925	46.10	0.00%	3.770
18	0.918	46.70	0.14%	4.234
19	0.910	47.79	0.30%	4.776
20	0.903	47.42	0.00%	5.275
21	0.910	47.92	0.87%	6.133
22	0.905	48.93	0.51%	6.577
23	0.872	50.22	0.30%	7.087

About 2.7 hours of SAE training across all 24 layers on a single H100 SXM, holding roughly 246k tokens/sec on the optimizer steps. That is training time only. Wall clock was longer because the run got restarted while I fixed two trainer bugs.

Sparsity holds everywhere. Every layer lands between 46.1 and 50.4 against a target of 50, and dead features stay under 1% on all twenty-four.

The floor begins at layer 4

Layers 0 through 3 do not decompose. 0.62 to 0.73 explained variance, and they burn 4,200 to 4,700 steps getting there. Layer 4 lands 0.990 in 1,751 steps.

Layer 4 is not remarkable because 0.990 is a big number. Layer 5 is 0.988 and layer 6 is 0.987. It is remarkable because of the size of the step to get there. Every other layer-to-layer transition in this model moves EV by less than 0.04. The L3 to L4 step moves it by 0.262.

That is a cliff, and it lines up with the model rather than the trainer. The residual RMS column sits between 0.133 and 0.327 for the first four layers, then jumps to 2.301 at layer 4. Seven times, across one layer boundary, exactly where the EV cliff is.

Below the cliff the stream is small and flat, lots of comparable-magnitude directions and nothing for fifty features to grab. Above it a handful of directions carry the energy and fifty features is plenty.

This model has a 130,560-token bilingual vocabulary, and untied embeddings account for about 401M of its 1,081M parameters. Roughly 37% of MiniCPM5-1B is vocabulary. Four layers of dense, hard-to-decompose stream.

EV tracks the residual norm all the way down

Past the cliff, the two columns move against each other for nineteen straight layers. Residual RMS climbs monotonically from 0.133 at layer 0 to 7.087 at layer 23, a 53x increase end to end, and explained variance walks down with it: 0.990 at layer 4, 0.954 at 12, 0.933 at 15, 0.903 at 20, 0.872 at 23.

Global-variance EV is easier to satisfy when a few directions dominate the stream, so as the stream gets louder and more concentrated with depth, fifty features cover progressively less of it. Two small ticks upward at layers 17 and 21, neither big enough to explain.

Hyper parameter adjustments aren't required

No hyperparameter adjustments across the whole run.

The scheduler drives itself. Sparsity is a constraint solved with augmented Lagrangian math, not a loss term you tune by hand and hope about. A dual variable rises until L0 sits where it was told to sit, a proportional controller nudges the JumpReLU threshold when the straight-through gradient goes blind, and the trainer determines without manual input when a layer is finished instead of burning the full 5,000-step budget.

It also picks which exit to use. The hard early layers ended on a sparsity-convergence timeout, the scheduler giving up on reconstruction once L0 had settled and stopped improving. The good layers ended on a quality gate instead: L0 inside 15% of target with EV above 0.95 held for 500 steps. Two different stop conditions.

Layer 0 started at an L0 probe of 73.0 and walked itself down to 48.9. I watched it.

Two bugs discovered in my trainer while working with this model

Layers 13 through 15 initially finished at L0 around 1,700 against a target of 50, which is not a K=50 SAE. What happens: once sparsity locks and the controller pins the sparsity pressure, the feature tail starves. Dead features went 0.8% to 8.8% in a single 250-step window while EV barely moved, 0.949 to 0.948. Then the emergency revival dumped thousands of features back in at once, and with the pressure term frozen there was nothing left to pull sparsity back down.

Two fixes. One releases the frozen pressure term when sparsity escapes its band. The other was a cached dead-feature set that was being cleared on 249 of every 250 steps, so the gentle revival path ran 0.4% of the time.

I added a hard 1% dead-feature ceiling with rollback: buffer recent states, and when dead breaches the ceiling, walk back to the last state under it and stop there. On layer 13 that cost one thousandth of EV for 8.6x fewer dead features.

Caveats

Layers 0 to 3 are weaker reconstructions. 0.62 to 0.73 EV on fifty features.

English-only corpus on a bilingual model. MiniCPM5-1B is trained on English and Chinese. The probe corpus was HuggingFaceFW/fineweb-edu, English only. So these features are what English text pulls out of a bilingual model, and the Chinese half is unmapped. The brain atlas for this model found that around 82% of probed features promote at least one CJK token, flat at every depth. This presents a known gap.

EV is a blunt ruler. Global-variance explained variance rewards streams with a few dominant directions, which is very likely part of what the residual RMS column is showing.

One corpus, one seed. FineWeb-Edu, seed 0.

The trainer is still under development and not published. The MiniCPM5-1B weights are from a pipeline you cannot yet audit. It found two of its own bugs on this model, which is the honest reason the deep-layer weights are shaped the way they are. I'm just a bartender. Use the artifacts with some caution.

Load one and check the sparsity yourself

No trainer install needed, this is plain torch against the state dict.

# is the constraint real, or did I just say it was?
import torch

sd = torch.load("layer_04_s0/sae.pt", map_location="cpu")
W_enc_w = sd["W_enc.weight"]        # [49152, 1536]
W_enc_b = sd["W_enc.bias"]          # [49152]
b_dec   = sd["b_dec"]               # [1536]
thr     = sd["log_threshold"].exp() # [49152]

acts = ...  # [tokens, 1536] residuals from MiniCPM5-1B layer 4

pre = (acts - b_dec) @ W_enc_w.T + W_enc_b   # decoder bias centers the input
feats = pre * (pre > thr)                    # JumpReLU, hard gate per feature

print(feats.shape[-1])                       # 49152 features
print((feats > 0).sum(-1).float().mean())    # should sit right around 50

Links

• Base model: https://huggingface.co/openbmb/MiniCPM5-1B
• Probe corpus: https://huggingface.co/datasets/HuggingFaceFW/fineweb-edu
• Brain atlas for the same model: https://huggingface.co/datasets/juiceb0xc0de/MiniCPM5-1B-atlas