You can deploy this model and start earning money today!

575-lab/
kiji-inspector-google-gemma-4-E2B-it

Name: kiji-inspector-google-gemma-4-E2B-it
Availability: InStock

Views

No views yet

kiji-inspector — SAE interpretability run: `google/gemma-4-E2B-it`

JumpReLU Sparse Autoencoders trained on residual-stream activations of google/gemma-4-E2B-it (35-layer decoder, d_model = 1536), to probe how the model selects tools across network depth.

Subject model: google/gemma-4-E2B-it — 35 layers, d_model = 1536.
Layer sweep: 13, 14, 15, 18, 19, 20, 23, 24, 25 (depths ≈ 37 % – 71 %, mid-to-deep).
SAE width: d_sae = 16384 (≈ 10.7× expansion over d_model).
Sparsity target: target_l0 = 75; 10 epochs; float32 activations.
Generation / labelling / judge model: Qwen/Qwen3-VL-235B-A22B-Instruct-FP8.

SAE health

Layer	Depth	d_model	Alive %	Dead %	L0 (mean)	Recon MSE
13	37 %	1536	5.03	58.54	95.31	0.1136
14	40 %	1536	4.32	74.69	655.08	20.5775
15	43 %	1536	2.99	74.85	482.61	3.4302
18	51 %	1536	1.28	71.66	173.75	0.1177
19	54 %	1536	0.91	68.80	115.18	0.0710
20	57 %	1536	1.03	70.61	145.38	0.0673
23	66 %	1536	1.21	75.39	194.35	0.0515
24	69 %	1536	0.98	72.97	155.20	0.0688
25	71 %	1536	1.04	68.88	162.99	0.0641

L0 vs target. No layer reached target_l0 = 75; L13 is closest (95). L14 and L15 are under-converged — L0 overshoots to 655 / 483 and reconstruction MSE blows up to 20.6 / 3.4 (vs ~0.05–0.12 elsewhere). Treat those two SAEs as not-yet-trained rather than as faithful decompositions. Deeper layers (18–25) reconstruct cleanly (MSE 0.05–0.12) but still carry L0 ≈ 115–194, i.e. 1.5–2.6× the sparsity target. Alive % is low (1–5 %) and dead % high (59–75 %), consistent with the aggressive 10.7× dictionary expansion.

Token-level fuzzing (autointerp)

Combined score = 0.7·token-accuracy + 0.3·prompt-accuracy. Quality tiers: Excellent > 0.8, Good 0.6–0.8, Poor < 0.6.

Layer	Depth	Features eval	Combined	Token-acc	Excellent %	Good %	Poor %
13	37 %	201	0.7054	0.6453	20.9	53.7	25.4
14	40 %	650	0.7222	0.6474	26.0	55.2	18.8
15	43 %	483	0.7146	0.6366	22.8	55.1	22.1
18	51 %	180	0.7153	0.6528	34.4	40.6	25.0
19	54 %	141	0.7391	0.6546	27.7	57.5	14.9
20	57 %	158	0.7141	0.6430	22.8	58.2	19.0
23	66 %	196	0.7061	0.6480	26.0	49.5	24.5
24	69 %	159	0.6859	0.6214	17.6	54.1	28.3
25	71 %	168	0.6952	0.6310	20.8	54.8	24.4

Best layer: L19 — highest combined (0.739) and token-acc (0.655), and the lowest poor rate (14.9 %). L18 has the most excellent features (34.4 %) but a higher poor rate (25 %). Fuzzing quality is fairly flat across the mid-to-deep sweep (combined 0.69–0.74); the two under-converged layers (14, 15) still fuzz reasonably because the evaluation scores feature descriptions, not reconstruction fidelity.

Per-layer artefacts

Each layer_<N>/ contains:

sae_checkpoints/sae_final.pt — trained JumpReLU SAE weights.
sae_checkpoints/config.json — training hyper-parameters.
sae_checkpoints/feature_health.json — alive/dead, L0, reconstruction MSE (with 95 % CIs).
sae_checkpoints/metrics.jsonl — per-step training log.
activations/contrastive_features.json — Cohen's-d tool-selection features (post-hoc probe).
activations/feature_descriptions.json — autointerp labels.
activations/fuzzing_summary.json / fuzzing_results.json — token-level fuzzing eval.
activations/decision_report.json — per-layer decision summary.
activations/metadata.json — model, layer, d_model, token/shard counts.

Citation

If you use these artefacts, please cite the kiji-inspector framework (Opening the Black Box: Mechanistic Interpretability of Agent Tool Selection with Sparse Autoencoders).

575-lab/kiji-inspector-google-gemma-4-E2B-it

kiji-inspector — SAE interpretability run: google/gemma-4-E2B-it