qwen3.8-27b-atlas

A brain atlas for Qwen/Qwen3.8-27B, a 64-layer hybrid linear/full-attention language model (the qwen3_5 architecture) that also carries a vision tower. This is not a chat dataset or a benchmark. It is an internal-mechanics map built by running activations through a corpus of prompts and scoring what each layer, component, head, and feature direction is doing.

The atlas is published as Parquet so it is queryable in the Hugging Face Dataset Viewer SQL Console. Every table below is a SQL view. The links in Query it on the Hub open the console with a real query preloaded. No install, no download, no compute.

---

Edit - August 26, 2026

27-31 accuracy band. Those numbers are exact 64ths: 0.9844 is 63/64, 0.9688 is 62/64, 0.9375 is 60/64. The whole "clean band" is three samples out of sixty-four, and every layer was already sitting near ceiling. That's the resolution floor of the probe, not structure in the model. I disproved the band.

Corporate-versus-authentic results. The two corpora are not minimal pairs. They differ in register, vocabulary, length and structure all at once, so the contrast separates two piles of text, not two behaviors. The tell is already printed above: 591 of 1280 heads flagged corporate-leaning. When 46% of the building is selective, nothing is. Disregard compliance_behaviour_features and compliance_behaviour_per_head.

I left the original text intact for any users who may have used this database and noticed the errors themselves.

---

Query it on the Hub

Each link opens the in-browser DuckDB SQL console on this dataset with the query already typed and run. Bookmark them, share them, edit the SQL in the URL.

Anomalous layers and surgery risk

• Sub-Zero classifier accuracy by layer (peak at layers 27-31)
• highest-damage surgery axes (the layer-0 landmine)
• Sub-Zero singular vectors

Compliance (corporate vs authentic contrast pass)

• compliance features, top contrast F-stat
• most authentic-leaning features (delta = mean_auth - mean_corp)
• most corporate-leaning features (delta = mean_auth - mean_corp)
• corporate-leaning heads (corp_leaning=1)

Features, heads, circuits

• feature taxonomy distribution (452 domain-specific out of 4.78M)
• domain-specific features only
• top F-stat features overall (late-layer gate)
• induction heads (OV circuits)
• per-head top F-stat
• strongest coactivations (q-component offset-32 pairs)
• logit-lens top features
• code-analysis roles (entangled vs selective)
• layer index, all 64 layers

If you want to know which directions carry a corporate-versus-authentic signal, which layers are dangerous to edit, where the induction heads live, or which feature directions move the logits most, this is the dataset. The queries above are the entry points.

What was run

• Model: Qwen/Qwen3.8-27B, qwen3_5 architecture, bf16 weights.
• Corpus: 8,965 prompts across 17 buckets (core_technical 550, ml_ai 550, research 527, writing 514, creative_writing 515, brainstorming 499, planning 472, data 467, learning 455, roleplay 589, introspection 589, business 590, design 590, niche 590, community 588, humor 590, tool_use 290). Raw census npz ~2.65 GB.
• Layers probed: all 64 text layers (0-63). 48 linear-attention layers and 16 full-attention layers (every 4th layer: 3, 7, 11, ..., 63).
• Passes: activation census and feature taxonomy, per-head statistics, OV/induction circuits, logit lens, coactivation, code-analysis, a corporate-versus-authentic compliance contrast pass, and Sub-Zero surgery scoring on all 64 layers.
• Atlas date: 2026-08-23.

Architecture notes

Property	Value
Architecture	Qwen3_5ForConditionalGeneration (VLM)
Text layers	64 (48 linear-attention + 16 full-attention)
full_attention_interval	4 (full attention every 4th layer)
hidden_size	5120
intermediate_size	17408
Attention (full layers)	24 query heads, 4 KV heads (GQA, group 6), head_dim 256
Linear attention	16 key heads x 128, 48 value heads x 128, conv kernel 4
MLP	SwiGLU (gate / up / down), width 17408
vocab_size	248320
RoPE	theta 1e7, partial rotary 0.25, mrope section [11, 11, 10]
tie_word_embeddings	false
Vision tower	27-layer ViT, hidden 1152, out 5120 (NOT probed)
MTP head	1 layer (NOT probed)

The source model is a vision-language model. This atlas probes only the 64 text layers. The 27-layer vision encoder and the multi-token-prediction head are outside the atlas scope, so any statement here is about the language path.

The atlas records the full-attention q projection as 12288 wide and k, v as 1024 each. With 24 query heads and 4 KV heads, that is a 6:1 GQA group; the q width is 24 x 512, double the config head_dim (256) times the head count, which is the Qwen3.5 query/key head-doubling convention. The heads component is 6144 (24 x 256, the attention output before the output projection) and attn is 5120 (back to hidden_size).

Linear-attention layers carry three components in the atlas: linattn_qkv (10240), linattn_z (6144, the 48 value heads x 128), and linattn_out (5120, back to hidden_size). The SwiGLU MLP (gate, up, mlp) appears in every layer at width 17408.

A geometry check passes exactly: 48 linear layers x 73728 coordinates + 16 full-attention layers x 77824 coordinates = 4,784,128, which equals the features row count. This confirms no component or layer was dropped in capture.

What the tables contain

Each table is one Parquet split and one SQL view in the console.

Table	Rows	What it gives you
features	4,784,128	Per-direction activation census: F-stat, activation rate, taxonomy class, per layer and component.
compliance_behaviour_features	4,784,128	Same directions scored on a corporate-versus-authentic contrast: fstat, delta (mean_auth - mean_corp), mean_corp, mean_auth.
coactivation	35,412	Strongest correlated feature pairs per layer/component, with dominant bucket.
logit_lens	17,408	Logit-lens F-stat per direction: which directions directly move the output distribution.
code_analysis	12,480	Per-direction code-analysis role (entangled vs selective, etc.).
subzero_capability	2,225	Sub-Zero surgery axes: damage (nats/token), domain, fence/freeze flags, explained variance.
per_head	1,280	Per-head best F-stat and specific-feature count (full-attention layers).
compliance_behaviour_per_head	1,280	Per-head compliance lean: corp_leaning flag, fstat_best, delta_at_top, top_dim.
subzero_svs	674	Sub-Zero singular vectors: classifier score, Wanda score, dark variance.
ov_circuits	384	OV circuit stats per full-attention head: induction score, spectral concentration (compliance score is NULL here).
subzero_layer	64	Per-layer Sub-Zero classifier accuracy and compliance behavior summary.
layers	64	Layer index and type (linear vs full attention).

The sae_features table is empty (no SAE pass was run), so it is omitted from the Parquet set.

Key findings

Each finding names its numbers and links to the query that produced them.

Layer 0 is the model's surgery landmine

Layer 0 linattn_out_proj axis 0 does 3.522 nats/token of damage to math, 3.466 to code, 3.355 to reasoning, 3.281 to factual, and 3.268 to multilingual. fence_passed is 0, frozen is 0, and explained is 0.999. It is the single highest-damage axis in the model by roughly an order of magnitude: the next highest is layer 52 linattn_in_proj_z axis 0 at 0.364. The first linear-attention output projection carries a direction the whole model depends on.

Query: highest-damage surgery axes

Sub-Zero classifier accuracy peaks in a clean band at layers 27-31

Layers 27, 28, 29, 30, and 31 all score 0.9844 classifier accuracy. Layer 0 and layers 13-18 sit at 0.9688. The lowest layers are 5-11 at 0.9375. The editability signal is strongest in the late-mid block, not at the top or bottom of the stack.

Query: classifier accuracy by layer

The compliance signal concentrates in the final MLP

The strongest corporate-versus-authentic directions by contrast F-stat live across layers 17-29 and across components (up, mlp, linattn_z, linattn_qkv, gate). The extreme leaners, though, pile into the last four MLP layers. The most authority-leaning direction is L63 mlp 7528 (delta +12.29, mean_auth 14.00 vs mean_corp 1.71). The most corporate-leaning is L60 mlp 269 (delta -22.98). A cleaner corporate example is L63 mlp 1764 (mean_corp 11.28, mean_auth 0.27): it fires high on corporate prompts and near-zero on authority. The final MLP is where the corporate/authority commitment gets written.

Query: top contrast F-stat / authentic-leaning / corporate-leaning

591 of 1280 heads are flagged corporate-leaning, concentrated in q at layers 19 and 23

The per-head compliance pass flags 591 heads corp_leaning=1 and 689 at 0. The strongest corporate-leaning heads are query heads in a narrow band: L19 q H29 (fstat_best 5187), L23 q H14 (4689), L23 q H38 (4603), L23 q H0 (4519), L31 q H26 (4303). The corporate read forms in the query projection of the mid full-attention layers.

Query: corporate-leaning heads

Induction heads cluster at layer 27 head 23

The top induction head is L27 H23 at induction score 1.192, followed by L7 H23 (1.128), L11 H15 (1.109), L7 H8 (1.107), and L27 H15 (1.105). Head 23 recurs at layers 7, 27, and 31. The OV compliance_score column is NULL for all 384 rows, so read this table for induction, not compliance.

Query: induction heads

452 domain-specific features in 4.78 million, and they punch far above their weight

Of 4,784,128 directions, 1,911,723 are non_activated and 1,400,306 are all_shared. Only 452 are domain-specific (specific_*), led by specific_creative_writing (118), specific_tool_use (73), specific_introspection (69), specific_learning (52), and specific_ml_ai (39). The specific features have a mean F-stat of 402.6 for creative writing against 42.6 for all_shared. Before reading the 452 as a finding about the model, read it as a finding about the instrument: a general-purpose corpus of 8965 prompts will undercount a model's specialized directions. A targeted corpus would surface more.

Query: taxonomy distribution / domain-specific only

The top of the F-stat table is all late-layer gate

The highest-F-stat directions in the model are all gate components in layers 50-63: L60 gate 3054 (fstat 958, activation_rate 1.0), L52 gate 2694 (904), L56 gate 12003 (818), L51 gate 3525 (800). The logit-lens top features are the same set, in the same order. The directions that most strongly select tokens at the output are the late-layer SwiGLU gates.

Query: top F-stat overall / logit lens

Coactivation in the q component has a clean offset-32 structure

The strongest coactivations are all in the q component and pair indices separated by 32: L51 q (10240, 10272) at correlation 1.0, L59 q (4610, 4640) at 0.9999 (dominant bucket business), L51 q (8224, 8192) at -0.9998. The offset 32 is a structural stride inside the 12288-wide query projection, and the pairs are near-perfectly correlated or anti-correlated. That is architecture-level redundancy in the query path, not a per-prompt coincidence.

Query: strongest coactivations

What Sub-Zero is measuring

Sub-Zero is a surgical-edit risk pass. For each candidate direction (a singular vector of a projection), it ablates that direction on a held-out probe and measures the log-likelihood damage in nats per token, across five domains: math, code, reasoning, factual, and multilingual. A high damage means the model depends on that direction. fence_passed records whether the direction cleared a safety fence, and frozen whether it was held out of surgery. explained is the variance the singular vector accounts for. The damage numbers are a relative map of where the model is fragile, not a license to edit: in this atlas fence_passed is 0 across every axis, so no direction was cleared.

Important caveats

• Text path only. The source model is a VLM with a 27-layer vision tower and an MTP head. This atlas probes the 64 text layers only. The vision encoder and MTP head are not represented.
• General-purpose corpus. The 8965-prompt, 17-bucket corpus is not specialized. The 452 domain-specific features are a lower bound on what the model represents, not a ceiling. A targeted corpus is the follow-up run that would raise that count.
• compliance_behaviour_features is a companion to features. It is the same 4,784,128 directions scored on the corporate-versus-authentic contrast. delta is mean_auth - mean_corp: positive means authentic-leaning, negative means corporate-leaning.
• ov_circuits.compliance_score is NULL for all 384 rows. The compliance score did not resolve on the OV pass. Read ov_circuits for induction and spectral concentration, not compliance.
• Sub-Zero damage is relative. fence_passed is 0 on every axis in this atlas. Treat the damage numbers as a comparative fragility map, not as clearance to ablate.
• Spectral and rank metrics are dimension-dependent. Effective rank, spectral concentration, and head counts are not comparable across model families without normalizing by head dimension. Use them within this atlas.
• No SAE pass. sae_features is empty and is omitted from the Parquet splits.

How to use

In the browser (no install). Open any link in Query it on the Hub. The SQL console runs DuckDB on the Parquet splits client-side. Edit the query in the URL or in the console.

DuckDB CLI with hf:// (no browser memory cap). The console is bounded by browser memory. For the 4.78M-row features split, use the DuckDB CLI and the Hugging Face community extension:

INSTALL huggingface;
LOAD huggingface;
-- each split is one parquet file under data/
SELECT layer_id, classifier_accuracy, compliance_behaviour_pct
FROM read_parquet('hf://datasets/juiceb0xc0de/qwen3.8-27b-atlas/data/subzero_layer.parquet')
ORDER BY classifier_accuracy DESC;

Local Python on the canonical SQLite. The Parquet is a derived query surface. The canonical artifact is the SQLite file shipped in this repo at qwen3.8-27b/atlas.sqlite:

import sqlite3, pandas as pd
# the sqlite is the source of truth; parquet is the Hub query surface
con = sqlite3.connect("qwen3.8-27b/atlas.sqlite")
df = pd.read_sql(
    "SELECT layer_id, component, feature_idx, fstat, taxonomy_class "
    "FROM features ORDER BY fstat DESC LIMIT 20", con)
print(df)

Source model license

The source model Qwen/Qwen3.8-27B is released under the Apache License 2.0. This atlas, as a derived measurement of that model's activations, is published under the same license.

Contact / more

• Source model: Qwen/Qwen3.8-27B
• Atlas pipeline: the qwip_atlas activation-census pipeline
• Author: juiceb0xc0de