juiceb0xc0de/tmax-2b-atlas

A brain atlas for allenai/tmax-2b, a hybrid SSM/Mamba/transformer language model. This is not a chat dataset or a benchmark — it is an internal-mechanics map of the model, built by running activations through a corpus of prompts and scoring what each layer, component, head, and feature direction is doing.

If you want to know where the model stores compliance style, which late-layer directions you can edit without breaking reasoning, or whether the hybrid attention heads are copy-paste circuits, this is the dataset.

What was run

• Model: allenai/tmax-2b
• Corpus: 8,965 diverse prompts
• Layers probed: all 24
• Attention layers: 3, 7, 11, 15, 19, 23
• Hybrid layers: 0–2, 4–6, 8–10, 12–14, 16–18, 20–22
• Passes: activation census, feature taxonomy, per-head analysis, OV-circuit SVD, logit lens, coactivation, code-analysis, compliance-behavior contrast, Sub-Zero surgery with capability fence

What the tables contain

Table	Rows	What it gives you
layers	24	layer metadata and completion flags
features	681,984	feature taxonomy + activation stats per (layer, component, feature_idx)
per_head	168	per-head selectivity on the 6 attention layers
ov_circuits	48	SVD over W_V @ W_O plus QK/FC spectral metrics
logit_lens	6,528	promoted/suppressed output tokens per feature
coactivation	14,201	feature-pair correlations
code_analysis	4,680	entangled vs selective role labels
compliance_behaviour_features	681,984	authentic-vs-corporate contrast per feature
compliance_behaviour_per_head	168	per-head compliance/behavior contrast
subzero_layer	24	classifier accuracy and SV summary per layer
subzero_svs	96	bouncer singular values per projection
subzero_capability	365	DAS-axis damage and capability-fence decisions

Key findings

• Attention is distributed, not memorized. OV spectral concentration is 0.060 with effective rank ~79.
• MLP/SSM gates dominate the logit lens. Top logit-lens peaks are gate features in layers 15 and 20, with F-stats over 590.
• The worst surgical surprise is early. Layer 1 linattn_in_proj_z axis 0 fails the capability fence across all five domains, with up to 0.60 damage to multilingual.
• Otherwise editable. 335 of 365 Sub-Zero axes pass the capability fence (91.8%). Average damage is 0.029.

What Sub-Zero is measuring

The Sub-Zero pass is not a generic "find all important directions" sweep. It specifically looks for directions that separate corporate style from authentic style, then uses DAS rotation and a capability fence to check whether removing those directions damages code, math, reasoning, factual, or multilingual ability. The rows in subzero_capability are domain-by-domain damage scores for those compliance/behavior candidate axes, not a census of every load-bearing direction in the model.

Important caveat

The hybrid SSM/Mamba layers contain components that do not tokenize language. I deliberately did not probe those components because the activations would just produce noise with no interpretable structure. This atlas covers the language-tokenizing components only. I am working on a method to capture whatever those hybrid layers are actually doing, but it is not included here yet.

How to use

import sqlite3
import pandas as pd

conn = sqlite3.connect("tmax-2b-atlas.sqlite")
df = pd.read_sql_query("SELECT * FROM features WHERE layer_id=15 AND component='gate' ORDER BY fstat DESC LIMIT 20", conn)

Or just browse the tables with any SQLite viewer.

Backend environment (convenience only)

If you want to run Qwen3.5-family or tmax-family models yourself, the backend/ folder contains the Dockerfile and GitHub Actions workflow I use to build a CUDA 12.8 / torch 2.7 image with prebuilt flash-attn, causal-conv1d, mamba-ssm, and flash-linear-attention wheels. It is provided as-is and is not the main attraction of this dataset.

License

MIT.

Contact / more

• Model: https://huggingface.co/allenai/tmax-2b
• Atlas code: https://github.com/JuiceB0xC0de/qwip_atlas (or point to your current repo)
• Follow: https://huggingface.co/juiceb0xc0de