Datasets:

juiceb0xc0de
/

tmax-2b-atlas

component stringclasses 11 values	feature_a int64 0 6.14k	feature_b int64 0 6.14k	correlation float64 -1 1	dominant_bucket stringclasses 17 values
mlp	4,866	2,158	-0.956053	null
mlp	2,260	2,158	-0.939458	null
mlp	2,260	4,866	0.938029	business
mlp	2,539	4,866	-0.922664	null
mlp	2,539	2,158	0.917939	null
mlp	3,475	4,866	0.913529	business
mlp	2,260	2,539	-0.905134	null
mlp	3,475	2,158	-0.900024	null
mlp	3,475	2,260	0.895779	business
mlp	6,104	4,866	-0.893083	null
mlp	2,584	3,841	0.884372	writing
mlp	6,104	2,260	-0.882761	null
mlp	3,475	2,539	-0.878479	null
mlp	1,315	2,452	-0.878364	niche
mlp	6,104	2,158	0.878234	null
mlp	2,539	1,717	0.865931	null
mlp	4,434	4,866	0.862056	business
mlp	4,434	2,158	-0.861656	null
mlp	6,104	2,539	0.860231	null
mlp	1,717	4,866	-0.856753	null
mlp	3,475	1,717	-0.852585	null
mlp	4,991	2,158	0.851261	null
mlp	4,991	4,866	-0.849447	null
mlp	6,104	3,475	-0.848864	null
mlp	2,260	1,717	-0.848848	null
mlp	1,717	2,158	0.847156	null
mlp	3,414	4,866	0.844987	business
mlp	4,434	2,260	0.843902	business
mlp	4,991	2,260	-0.838803	null
mlp	1,937	2,158	0.832946	null
mlp	4,434	1,717	-0.83078	null
mlp	3,414	2,158	-0.830372	null
mlp	792	4,866	-0.830296	null
mlp	5,309	4,866	0.830162	business
mlp	4,991	1,717	0.827987	null
mlp	4,434	3,475	0.827822	business
mlp	792	2,158	0.826548	null
mlp	3,414	2,539	-0.826361	null
mlp	5,795	4,568	-0.821368	learning
mlp	4,434	2,539	-0.819726	null
mlp	491	4,568	-0.818553	business
mlp	5,309	2,158	-0.815355	null
mlp	6,104	1,717	0.811764	null
mlp	4,991	2,539	0.810573	null
mlp	491	5,795	0.808733	roleplay
mlp	5,309	2,260	0.808326	business
mlp	4,991	3,475	-0.803397	null
mlp	3,414	2,260	0.802777	business
mlp	1,937	4,866	-0.80209	null
mlp	3,414	3,475	0.798431	business
mlp	5,309	2,539	-0.795626	null
mlp	2,539	1,937	0.7924	null
mlp	4,434	4,991	-0.792222	null
mlp	792	2,260	-0.790679	null
mlp	3,414	1,717	-0.789507	null
mlp	2,260	1,937	-0.7874	null
mlp	1,112	1,751	-0.786942	humor
mlp	4,991	6,104	0.785197	null
mlp	5,309	3,475	0.783801	business
mlp	4,434	6,104	-0.78376	null
mlp	792	3,475	-0.773429	null
mlp	1,555	2,158	0.772284	null
mlp	5,309	6,104	-0.771935	null
mlp	6,027	4,866	0.771638	business
mlp	1,555	4,866	-0.771227	null
mlp	4,568	3,580	0.770878	niche
mlp	792	2,539	0.770576	null
mlp	3,414	6,104	-0.770041	null
mlp	4,434	3,414	0.769463	business
mlp	1,555	2,539	0.762936	null
mlp	6,027	2,158	-0.761703	null
mlp	5,309	4,434	0.760103	business
mlp	792	6,104	0.756937	null
mlp	2,260	6,027	0.754545	business
mlp	3,475	6,027	0.754259	business
mlp	4,685	2,350	-0.751068	roleplay
mlp	5,442	2,952	-0.744884	niche
mlp	6,104	1,937	0.742812	null
mlp	792	1,937	0.74262	null
mlp	1,555	2,260	-0.739766	null
mlp	4,991	6,027	-0.737755	null
mlp	5,309	1,717	-0.736511	null
mlp	6,027	1,717	-0.733089	null
mlp	3,885	207	-0.732932	humor
mlp	5,309	3,414	0.73244	business
mlp	5,309	4,991	-0.731905	null
mlp	1,284	1,966	-0.730243	research
mlp	6,027	2,539	-0.729061	null
mlp	5,421	4,866	-0.727464	design
mlp	491	3,580	-0.725593	business
mlp	5,309	792	-0.723716	null
mlp	3,414	792	-0.722403	null
mlp	5,421	2,539	0.722096	null
mlp	6,043	2,350	0.72118	roleplay
mlp	5,421	3,475	-0.720114	design
mlp	3,475	1,937	-0.718261	null
mlp	1,555	6,104	0.716792	null
mlp	5,421	1,717	0.716206	null
mlp	1,717	6,069	-0.71493	null
mlp	5,421	2,260	-0.714531	design

End of preview. Expand in Data Studio

`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.