juiceb0xc0de/smollm2-135m-instruct-atlas-SAE

SmolLM2-135M-Instruct SAE Atlas

30 custom SAEs across all 30 layers of HuggingFaceTB/SmolLM2-135M-Instruct. Zero dead features at convergence. Mean explained variance 0.9519. 552,960 sparse features. One sitting, one trainer, zero hyperparameter tuning.

What is this

A full mechanistic interpretability atlas for SmolLM2-135M-Instruct — every layer, every component (MLP, gate, up-projection, attention heads, Q/K/V), with SAE sparse features scored and merged into a 91.6MB SQLite database. The SAEs were trained with a custom trainer built from the ground up for this run. No OpenAI or Eleuther SAE for guidance on where to start.

The census corpus that drove it: 4,946 prompts across 16 buckets — technical, creative, reasoning, roleplay, introspection, humor, niche, the works. Not random web scrap. Curated to light up different circuits.

Results

Layer	EV	Mean L0	Recon Loss	Dead %
0	0.9480	48.74	0.2074	0.0
1	0.9599	43.65	0.3298	0.0
2	0.9631	46.81	0.5021	0.0
3	0.9508	46.56	0.7462	0.0
4	0.9463	46.23	0.8936	0.0
5	0.9350	47.57	1.1605	0.0
6	0.9306	48.44	1.3838	0.0
7	0.9318	49.51	1.5446	0.0
8	0.9432	46.52	1.6598	0.0
9	0.9373	47.15	2.0706	0.0
10	0.9348	45.53	2.2983	0.0
11	0.9905	48.58	5.8113	0.0
12	0.9901	48.42	6.1039	0.0
13	0.9891	46.15	6.9692	0.0
14	0.9884	44.76	7.1844	0.0
15	0.9863	47.63	8.6521	0.0
16	0.9840	43.45	10.179	0.0
17	0.9808	45.88	12.363	0.0
18	0.9811	47.17	12.300	0.0
19	0.9775	46.23	15.521	0.0
20	0.9726	48.23	18.727	0.0
21	0.9667	46.49	24.817	0.0
22	0.9594	46.15	31.128	0.0
23	0.9418	45.13	47.936	0.0
24	0.9397	45.35	57.931	0.0
25	0.9299	46.33	74.413	0.0
26	0.9212	45.32	92.727	0.0
27	0.9139	45.69	118.75	0.0
28	0.8809	46.56	129.83	0.0
29	0.8812	52.01	196.36	0.0

Highlights:

• Dead features: 0.0% across all 30 layers. Not averaged. Every single layer.
• Mid layers 11–18: EV above 0.98, peaking at 0.9905 on layer 11
• L0 stayed tight: 43.45–52.01 against a target of 50
• Mean EV across all layers: 0.9519

SAE training

Target L0 of 50, rolling-HF capture, 500 pool batches, 5000 max steps, 32,768 tokens per microbatch. Custom trainer — built from scratch for this run, not adapted from an existing library. The full training log and per-layer breakdown live in smollm2-135-instruct-SAE-trainer.md. The raw Modal terminal output is preserved in smollm2-135-instruct-SAE-trainer.raw.log.

Layer 29 needed a separate pass due to an off-by-one in the layer range argument — 0,29 trained layers 0–28, not 0–29. It still converged clean — EV 0.8812, L0 52.01, 0 dead features.

What's in here

If you know the layer, go to atlas/layers/<layer>/. That's the home base. Everything else branches from there.

atlas/

• atlas.sqlite — the main database (91.6MB)
• manifest.json — layer list, components, corpus metadata
• model_meta.json — model reference
• layers/<layer>/ — per-layer research artifacts:
  • heatmaps (gate, attention, heads, Q/K/V)
  • separation scores
  • code analysis
  • coactivation pairs
  • neuron taxonomies (MLP, gate, up, attention, Q/K/V)
  • per-head summaries
  • component comparison
  • compliance_behaviour/ — compliance/behaviour scoring
  • sub_zero/ — Sub-Zero specific outputs
  • census/ — per-layer census detail

raw/

• 30 raw census .npz files (l<layer>_census_raw.npz)
• census_raw_index.parquet — index into the census dumps

SAE/

• 30 SAE score .npz files (sae_l<layer>_custom_l0_50.npz)
• sae_features.parquet — 552,960 rows of sparse feature data
• sae_files.parquet — file-level SAE metadata

exports/

• brain_atlas_layers.parquet — viewer-friendly layer summary (552KB)
• brain_atlas_summary.json — aggregate atlas summary
• compliance_behaviour_scores.parquet — 201,600 rows of compliance/behaviour scores
• compliance_behaviour_scores.json — same data in JSON (56MB)

Quick lookup

You want	Go to
Layer 12 gate heatmap	atlas/layers/12/gate_census_heatmap.png
Layer 4 attention coactivation	atlas/layers/4/attn_coactivation_pairs.json
Layer 29 SAE scores	SAE/sae_l29_custom_l0_50.npz
Raw activations for layer 7	raw/l7_census_raw.npz
All compliance/behaviour scores	exports/compliance_behaviour_scores.parquet
Layer-level summary	exports/brain_atlas_layers.parquet

Corpus

The census was built from 4,946 prompts across 16 buckets:

core_technical · ml_ai · research · writing · creative_writing · brainstorming · planning · data · learning · roleplay · introspection · business · design · niche · community · humor

~300 prompts per bucket, designed to activate different circuits across the network. Not a random sample — targeted coverage.

Notes

• ov_circuits and subzero_svs are in the SQLite schema but empty in this release.
• SAE stays capitalized. Everything else lowercase.
• The SAE trainer code will be published separately. This dataset is the output.

How to query

The atlas is a SQLite database. If you're comfortable with numpy and parquet but haven't touched SQL, here's everything you need.

import sqlite3
import pandas as pd

db = sqlite3.connect("atlas/atlas.sqlite")

Get all features for a specific layer and component:

df = pd.read_sql("""
    SELECT * FROM features
    WHERE layer_id = 12 AND component = 'gate'
""", db)

Find the strongest compliance/behaviour features:

df = pd.read_sql("""
    SELECT * FROM compliance_behaviour_features
    WHERE fstat > 100
    ORDER BY fstat DESC
""", db)

Get SAE features sorted by topic strength:

df = pd.read_sql("""
    SELECT * FROM sae_features
    WHERE layer = 12 AND variant = 'custom_l0_50'
    ORDER BY topic_fstat DESC
    LIMIT 20
""", db)

Which features are specific to a circuit vs broadly shared?

df = pd.read_sql("""
    SELECT taxonomy_class, COUNT(*) as n
    FROM features
    WHERE layer_id = 12
    GROUP BY taxonomy_class
    ORDER BY n DESC
""", db)

Coactivation pairs — what fires together:

df = pd.read_sql("""
    SELECT * FROM coactivation
    WHERE layer_id = 12 AND component = 'mlp'
    ORDER BY ABS(correlation) DESC
    LIMIT 20
""", db)

Per-head attention analysis:

df = pd.read_sql("""
    SELECT * FROM per_head
    WHERE layer_id = 12 AND component = 'heads'
    ORDER BY fstat_best DESC
""", db)

Tables at a glance:

Table	Rows	What it holds
features	201,600	Per-layer, per-component feature taxonomies + activation stats
sae_features	394,473	SAE sparse features with topic + compliance-behaviour fstat scores
coactivation	18,023	Feature pairs that fire together, with correlation + dominant bucket
per_head	720	Attention-head level: dims, specificity, fstat scores
code_analysis	7110	Feature roles: entangled or selective
compliance_behaviour_features	201600	Features scored on compliance vs. behaviour axis (fstat, delta, means)
layers	30	Per-layer metadata: prompts, corpus hash, capture timestamp
subzero_layer	30	Sub-Zero classifier accuracy, refusal angles, SV counts
subzero_svs	—	Individual singular vectors with classifier + wanda scores

Don't want SQL? The parquet exports in exports/ and SAE/ cover the big ones — load them with pd.read_parquet() and skip the database entirely.

Citation

If you use this atlas, cite it and the model:

@misc{smollm2-135m-sae-atlas-2026,
  author = {Holmberg, Rick},
  title = {SmolLM2-135M-Instruct SAE Atlas: Full-Layer Sparse Autoencoder Features},
  year = {2025},
  url = {https://huggingface.co/datasets/juiceb0xc0de/smollm2-135m-instruct-atlas-SAE}
}

@misc{smollm2,
  title = {SmolLM2},
  author = {{HuggingFace TB}},
  url = {https://huggingface.co/HuggingFaceTB/SmolLM2-135M-Instruct}
}