juiceb0xc0de/qwen3.5-0.8b-cgs-atlas

A brain atlas for Qwen/Qwen3.5-0.8B, a 873M-parameter Qwen3_5ForConditionalGeneration with 24 layers, fine-tuned from Qwen/Qwen3.5-0.8B-Base. This is not a chat dataset or a benchmark. It is an internal-mechanics map, built by running 8,965 prompts through the model and scoring what every stream, head and neuron in all 24 layers does with them.

Every layer of the model answered the same set of questions: what does this prompt want, where in the stack is that answer readable, how wide is each layer's working space, and which neurons care about which kind of prompt. If you want to know where your model sorts its inputs by what they are for, which components carry that signal and which are dead weight, or where to look first before you consider editing, steering or ablating, this is the dataset.

Cartogrpahics Are you trainng models? Bring your own fine-tune and build an atlas. Find out what's happening under the hood of your own creations!

Headline numbers

Computed from the tables in this repo. What they mean is up to you. What they are is below.

• Most label-selective stream: up at layer 22, held-out probe accuracy 0.749 against a shuffled-label control of 0.063 (chance is 0.066). The probe gets about three prompts in four right.
• Neurons with a strong single-label lean: 26,995 neurons at |AUROC - 0.5| of 0.25 or more. Sort features by neg_log10_q to meet them.
• Widest working space: geometry participation ratio peaks at 98.50 in the mlp stream at layer 1; the tightest stream is attn_gate at layer 19 at 7.12.
• Dead surface: 0.00% of all captured neurons never move on any prompt in the corpus.
• Category readout across depth: label agreement of a prompt's 10 nearest neighbours runs from 0.375 in the early stack to 0.486 at layer 20, then settles to 0.426 by the last layer, 23.

What was run

• Model: Qwen/Qwen3.5-0.8B, revision 2fc06364715b967f1860aea9cf38778875588b17
• Corpus: 8,965 prompts across 17 usage buckets, 14 to 62 tokens, median 28, chat template applied
• Layers probed: all 24
• Streams: gate, linattn_out, linattn_qkv, linattn_z, mlp, up, attn, attn_gate, heads, k, q, v, each read at last_input (the final input token) and mean_all_real (mean over the prompt's tokens)
• Passes: sanity, geometry, held-out probes with a shuffled-label control, per-neuron ANOVA and taxonomy, per-label directions, coactivation, learned dictionaries, per-head norm analysis, cross-stream CKA, layer-by-layer CKA, neighbours, trajectories
• Not in this atlas: OV-circuit SVD, logit lens, contrast datasets, Sub-Zero surgery, code analysis

What you get

Table	Rows	What it gives you
coactivation	8,100	the 50 strongest correlated pairs among the 256 highest-F neurons, per stream per layer
dictionary_features	1,536	128 learned atoms per stream at its most selective layer: frequency, activation, top label, AUROC
features	473,088	one row per (layer, stream, neuron): ANOVA F, q, eta2, top label, AUROC, taxonomy
layers	336	one row per (layer, stream, position): sanity, geometry, probe, taxonomy counts, cross-stream CKA, neighbours, trajectories
per_head	168	per-head output-norm F, q and top label for heads, q, k and v
prompts	8,965	the corpus with labels, token counts, and per-stream outlier, hub and Mahalanobis scores

atlas/atlas.sqlite holds the same six tables. Every parquet config also opens directly in the Dataset Viewer's SQL console on this page: pick a config, hit SQL, and query without downloading anything.

from datasets import load_dataset

repo = "juiceb0xc0de/qwen3.5-0.8b-cgs-atlas"
features = load_dataset(repo, "features", split="train")
layers = load_dataset(repo, "layers", split="train")

Three things to try in five minutes

1. Find the neurons that split a label from everyone else. In the SQL console on the features config: SELECT layer, component, neuron, top_label, top_auroc FROM features WHERE top_label = 'roleplay' AND q < 0.01 ORDER BY abs(top_auroc - 0.5) DESC LIMIT 20. Swap roleplay for any label in prompts. AUROC 1 means every prompt of that label sits above every other prompt on that neuron; a value under 0.5 is an equally strong negative lean.
2. Watch the working space across depth. On the layers config: SELECT layer, component, geometry_participation_ratio, geometry_twonn FROM layers WHERE position = 'mean_all_real' AND component = 'up' ORDER BY layer. Participation ratio is an effective dimension; a value near 1 means one direction holds the whole stream.
3. Diff it against the base. If this model is a fine-tune and the base's atlas is published, run the same two queries against both. What the tune kept and what it moved is exactly what an editor or ablator wants to know before touching a weight. The base, Qwen/Qwen3.5-0.8B-Base, has no public atlas yet; point the Space at it and it will.

Reading the numbers

• Positions. last_input: activation at the final input token; with a chat template this is the last token of the generation prompt. mean_all_real: mean over every real (unpadded) token, template and special tokens included. Per-neuron and per-prompt numbers use the mean position.
• AUROC below 0.5 is not weak. It is a negative association as strong as the same distance above 0.5: the label's rows run low on that neuron. The association column says which way.
• Rank by neg_log10_q, threshold on q. q is float64 and reads 0 once the corrected p-value is below about 1e-308; neg_log10_q stays finite and keeps neurons ordered. Significant is q below 0.01.
• Held-out means held out. The PCA and the score scaling are fit on each training fold and applied to its held-out fold, so no held-out prompt shapes its own coordinates. The control probe runs on shuffled labels through the same folds.
• Labels are usage categories, not capabilities. Some are sub-genres of others and read alike in wording and theme. Two prompts that are near copies can carry different labels; the probe counts those as errors.
• Dominated streams. When the participation ratio is below 3, one direction holds the whole stream and the geometry, trajectory and neighbour numbers describe that direction and nothing else. The probe is unaffected. These rows are flagged in layers as geometry_dominated and listed in cross_layer/geometry.json.
• Short prompts. This corpus runs 14 to 62 tokens. The first principal component of early layers tracks prompt length; pc_length_corr in each layer's summary.json gives the correlation.
• Descriptive only. Every number here was measured by watching, not by intervening on the model.

What an atlas is, and what this one is not

An atlas maps how the model reacts to prompts. This is a census run: activations read at two token positions inside the input, scored against 17 usage labels. It does not capture the model speaking. What the model would generate back, and where its internals diverge while it generates, is a different measurement and a different atlas.

Every measurement in this repo is a standard, named technique: one-way ANOVA with Benjamini-Hochberg correction, held-out multinomial probes over principal components, linear CKA, TwoNN intrinsic dimension, sparse dictionary learning, kNN and k-means label agreement. The full list is in the schema below, and every number traces to it. What is not public is the orchestration that runs all of it together on commodity hardware in minutes instead of hours. That part is the product, and it is how the atlases stay free.

Provenance

• model: Qwen/Qwen3.5-0.8B, revision 2fc06364715b967f1860aea9cf38778875588b17
• weights commit: 2fc06364715b967f1860aea9cf38778875588b17
• capture format census, version 2, fingerprint sha256:6c87bca7bf6ca31765d39d4d849e3b39dc7bc8c660c439c9c6cc96315aa6075a
• capture code 88a0f8e40f064d1b92b44f2fe7ef56a98ed8a9d9-dirty
• analysis code 203ff7bddba7429fb3bf38e1ac48ccf54a341ba9
• capture libraries: causal_conv1d 1.7.0, flash_linear_attention 0.5.2, numpy 1.26.4, python 3.10.13, torch 2.13.0+cu130, transformers 5.17.0, triton 3.7.1
• profile calibration
• job qwen3.5-0.8b-census-261022-024
• 8,965 prompts, 17 labels, 24 layers, streams gate, linattn_out, linattn_qkv, linattn_z, mlp, up, attn, attn_gate, heads, k, q, v at positions last_input, mean_all_real
• prompt length 14 to 62 tokens, median 28

Want your model mapped?

This atlas was curated by the cartographics atlas builder. You can get the same treatment for any public causal LM on the Hub: open juiceb0xc0de/cartographics, point it at your model, and let it run. The capture burns a few minutes of your own ZeroGPU quota; everything after that, the census, the statistics and the assembly, runs on the backend here. The finished atlas lands as a dataset repo like this one.

Every atlas in the series lives under juiceb0xc0de. If your end goal is editing, steering or ablation, the pair you want is this atlas plus the base model's, and the same two queries on both.

License

The source model is released under apache-2.0, and that license applies to the weights the atlas measures; see Qwen/Qwen3.5-0.8B.

Contact

Questions, corrections, or a model you want prioritized: open a discussion here or on the cartographics Space. Atlas pipeline and methodology posts: huggingface.co/juiceb0xc0de.

---

Schema reference

Per-column definitions for every table, kept close to the capture's own documentation. The file-by-file map of atlas/ is in atlas/README.md.

coactivation

• layer: layer index, 0 is the first block's output
• component: the stream: resid, mlp, attn, heads, gate, up
• a: neuron index, one of the 256 neurons with the highest F in this stream at this layer
• b: the other neuron of the pair
• r: Pearson correlation of the two neurons across prompts; the 50 pairs with the largest |r| are kept. For head streams, two neurons of one head are never paired

dictionary_features

• layer: layer index, 0 is the first block's output
• component: the stream: resid, mlp, attn, heads, gate, up
• atom: atom index, 0 to 127. The dictionary is learned at one layer per component: the layer where the probe's selectivity is highest
• frequency: share of prompts on which the atom is active
• mean_activation: mean |coefficient| over the prompts where the atom is active
• top_label: the label whose rows the atom's |coefficient| separates best from the rest
• top_label_auroc: one-vs-rest AUROC of the atom's |coefficient| for top_label
• dead: true when the atom is never active
• dense: true when the atom is active on more than half of the prompts

features

• layer: layer index, 0 is the first block's output
• component: the stream: resid, mlp, attn, heads, gate, up
• neuron: index into the component's feature vector; heads are flattened head by head
• F: one-way ANOVA F of the neuron across labels: variance between label means over variance within labels
• q: Benjamini-Hochberg false discovery rate of F among this stream's neurons at this layer, float64; significant below 0.01. Reads 0 once the corrected p is below about 1e-308
• neg_log10_q: -log10 of q from a log-space tail of the F distribution, finite for every finite F, so neurons stay ordered where q reads 0. Rank by this. inf means every label's rows share one value
• eta2: share of the neuron's variance explained by the labels, 0 to 1
• top_label: the label whose rows differ most from everyone else on this neuron, by |AUROC - 0.5|. With exactly two labels both lean the same amount and the first label is listed
• top_auroc: one-vs-rest AUROC of this neuron for top_label: 0.5 no separation, 1 every row of the label is above every other row, 0 every row is below
• association: positive when top_auroc is above 0.5 (the label's rows run high), negative below (they run low), none at exactly 0.5
• taxonomy: dead: never moves. inert: q at or above 0.01. selective: q below 0.01 and |AUROC - 0.5| at or above 0.25 for exactly one label. mixed: for two or more labels. broad: q below 0.01 but no label reaches 0.25

layers

• layer: layer index, 0 is the first block's output
• component: the stream: resid, mlp, attn, heads, gate, up
• position: which token position the stream was read at: mean_all_real (mean over the prompt's tokens) or last_input (the last token)
• sanity_rows: prompts in the stream
• sanity_features: width of the stream
• sanity_nan: NaN values found
• sanity_inf: infinite values found
• sanity_norm_p05: 5th percentile of row L2 norm
• sanity_norm_p50: median row L2 norm
• sanity_norm_p95: 95th percentile of row L2 norm
• sanity_linf_p50: median over rows of the row's largest |value|
• sanity_row_skew_p50: skew of a row's values across its features, median row
• sanity_row_kurtosis_p50: excess kurtosis of a row's values across its features, median row
• sanity_frac_abs_lt_1e-3: share of all values with |x| below 0.001
• sanity_frac_abs_lt_1e-2: share of all values with |x| below 0.01
• sanity_heavy_tailed_features: features whose excess kurtosis over prompts is above 20
• sanity_dead_features: features whose largest |value| over prompts is below 1e-6
• geometry_total_var: sum of per-feature variance
• geometry_anisotropy: mean cosine between distinct rows; 1 means every row points the same way
• geometry_mean_norm: L2 norm of the mean row
• geometry_anisotropy_centered: anisotropy after subtracting the mean row
• geometry_participation_ratio: (sum of eigenvalues)^2 over sum of squared eigenvalues of the covariance: an effective dimension. Spectrum from a 2048-row subsample
• geometry_entropy_rank: exp of the entropy of the normalized covariance spectrum: another effective dimension
• geometry_decay_alpha: slope of log eigenvalue against log rank over the top 100 eigenvalues
• geometry_k50: principal components needed to reach 50% of the variance
• geometry_k90: principal components needed to reach 90% of the variance
• geometry_k95: principal components needed to reach 95% of the variance
• geometry_k99: principal components needed to reach 99% of the variance
• geometry_twonn: TwoNN intrinsic dimension (Facco et al. 2017) on a 2000-row subsample
• geometry_dominated: true when the participation ratio is below 3: one direction holds the stream, and its geometry, trajectory and neighbor numbers describe that direction only
• geometry_pca_explained: share of the variance in the top 256 principal components
• trajectories_mahalanobis_p50: median Mahalanobis distance of a prompt from the layer's mean in the whitened top-256 PCA space
• trajectories_mahalanobis_p99: 99th percentile of that distance
• trajectories_cos_to_mean_p50: median cosine between a row and the layer's mean row
• layers_xcka_cka_with_gate.mean_all_real: linear CKA between this stream and the named stream at the same layer, on the same 2048 prompts
• layers_xcka_cka_with_linattn_out.mean_all_real: linear CKA between this stream and the named stream at the same layer, on the same 2048 prompts
• layers_xcka_cka_with_linattn_qkv.mean_all_real: linear CKA between this stream and the named stream at the same layer, on the same 2048 prompts
• layers_xcka_cka_with_linattn_z.mean_all_real: linear CKA between this stream and the named stream at the same layer, on the same 2048 prompts
• layers_xcka_cka_with_mlp.mean_all_real: linear CKA between this stream and the named stream at the same layer, on the same 2048 prompts
• layers_xcka_cka_with_up.mean_all_real: linear CKA between this stream and the named stream at the same layer, on the same 2048 prompts
• neighbors_knn_label_agreement: share of a prompt's 10 nearest prompts (cosine, top-256 PCA space) that carry its label, averaged over prompts
• neighbors_hubness_skew: skew of how many prompts list each prompt among their 10 nearest; high means a few prompts are everyone's neighbor
• neighbors_knn_cos_mean: mean cosine to the 10 nearest prompts
• neighbors_duplicates: prompt pairs with cosine at or above 0.98 in the top-256 PCA space
• neighbors_kmeans_ari: adjusted Rand index of k-means (one cluster per label) against the labels
• neighbors_kmeans_nmi: normalized mutual information of the same k-means against the labels
• neighbors_silhouette_labels: silhouette of the labels under cosine, on a 2000-row subsample
• probes_probe_accuracy: held-out accuracy of a multinomial logistic regression on the top-256 principal components, 5 stratified folds. The PCA and the score scaling are fit on each training fold and applied to its held-out fold
• probes_probe_accuracy_std: standard deviation of accuracy across the folds
• probes_probe_chance: share of the largest label: the accuracy of always guessing it
• probes_probe_pcs: principal components the probe used; fewer than 256 when the training fold is narrower
• probes_probe_folds: number of folds
• probes_probe_control_accuracy: the same probe on shuffled labels, same folds, same per-fold PCA
• probes_probe_selectivity: accuracy minus control_accuracy
• probes_median_F: median F over the stream's neurons
• probes_max_F: largest F
• probes_frac_significant: share of neurons with q below 0.01
• probes_eta2_mean: mean eta2 over neurons
• probes_taxonomy_dead: count of neurons in that taxonomy class (dead, inert, selective, mixed, broad)
• probes_taxonomy_inert: count of neurons in that taxonomy class (dead, inert, selective, mixed, broad)
• probes_taxonomy_selective: count of neurons in that taxonomy class (dead, inert, selective, mixed, broad)
• probes_taxonomy_mixed: count of neurons in that taxonomy class (dead, inert, selective, mixed, broad)
• probes_taxonomy_broad: count of neurons in that taxonomy class (dead, inert, selective, mixed, broad)
• features_coact_abs_r_mean: mean |correlation| among the 256 highest-F neurons; for head streams, pairs inside one head are left out
• features_coact_pairs_over_0.5: pairs among those 256 with |correlation| above 0.5; for head streams, pairs inside one head are left out
• trajectories_cos_prev_mean: mean cosine between a prompt's state here and one layer earlier
• trajectories_cos_prev_p05: 5th percentile of that cosine
• trajectories_rel_update_mean: mean of |x_l - x_(l-1)| / |x_(l-1)|: how much the layer moved each prompt, relative to where it was
• neighbors_knn_jaccard_prev: Jaccard overlap between a prompt's 10 nearest here and one layer earlier, averaged over prompts
• layers_xcka_cka_with_attn.mean_all_real: linear CKA between this stream and the named stream at the same layer, on the same 2048 prompts
• layers_xcka_cka_with_attn_gate.mean_all_real: linear CKA between this stream and the named stream at the same layer, on the same 2048 prompts
• layers_xcka_cka_with_heads.mean_all_real: linear CKA between this stream and the named stream at the same layer, on the same 2048 prompts
• layers_xcka_cka_with_k.mean_all_real: linear CKA between this stream and the named stream at the same layer, on the same 2048 prompts
• layers_xcka_cka_with_q.mean_all_real: linear CKA between this stream and the named stream at the same layer, on the same 2048 prompts
• layers_xcka_cka_with_v.mean_all_real: linear CKA between this stream and the named stream at the same layer, on the same 2048 prompts

per_head

• layer: layer index, 0 is the first block's output
• component: the stream: resid, mlp, attn, heads, gate, up
• head: head index within the component
• norm_F: one-way ANOVA F of the head's output norm across labels
• norm_q: Benjamini-Hochberg q of norm_F among the layer's heads; significant below 0.01
• norm_neg_log10_q: -log10 of norm_q in log space; rank by this
• norm_eta2: share of the head's output-norm variance explained by the labels
• norm_top_label: the label whose mean output norm sits furthest from the overall mean
• dim_F_mean: mean F over the head's dimensions
• dim_F_max: largest F among the head's dimensions
• dim_frac_significant: share of the head's dimensions with q below 0.01

prompts

• row_id: 0-based row in the dataset's corpus file
• label: the row's label, the class every per-label number uses
• category: top-level category from the corpus file, when the corpus has one
• subcategory: second-level category from the corpus file, when the corpus has one
• fine_bucket: finest grouping from the corpus file, when the corpus has one
• prompt: the prompt text, when the capture kept it
• n_tokens: prompt length in tokens as the model's tokenizer counts it, after any cut
• truncated: true when the prompt was cut to the capture's maximum length (128 tokens) before the forward pass; every number for that row saw only the head of the prompt
• attn_knn_outlier: 1 minus the mean cosine to the prompt's 10 nearest prompts, averaged over layers; high means isolated. Prefix is the component
• attn_hub_count: how many prompts list this one among their 10 nearest at the last layer. Prefix is the component
• attn_gate_knn_outlier: 1 minus the mean cosine to the prompt's 10 nearest prompts, averaged over layers; high means isolated. Prefix is the component
• attn_gate_hub_count: how many prompts list this one among their 10 nearest at the last layer. Prefix is the component
• gate_knn_outlier: 1 minus the mean cosine to the prompt's 10 nearest prompts, averaged over layers; high means isolated. Prefix is the component
• gate_hub_count: how many prompts list this one among their 10 nearest at the last layer. Prefix is the component
• heads_knn_outlier: 1 minus the mean cosine to the prompt's 10 nearest prompts, averaged over layers; high means isolated. Prefix is the component
• heads_hub_count: how many prompts list this one among their 10 nearest at the last layer. Prefix is the component
• k_knn_outlier: 1 minus the mean cosine to the prompt's 10 nearest prompts, averaged over layers; high means isolated. Prefix is the component
• k_hub_count: how many prompts list this one among their 10 nearest at the last layer. Prefix is the component
• linattn_out_knn_outlier: 1 minus the mean cosine to the prompt's 10 nearest prompts, averaged over layers; high means isolated. Prefix is the component
• linattn_out_hub_count: how many prompts list this one among their 10 nearest at the last layer. Prefix is the component
• linattn_qkv_knn_outlier: 1 minus the mean cosine to the prompt's 10 nearest prompts, averaged over layers; high means isolated. Prefix is the component
• linattn_qkv_hub_count: how many prompts list this one among their 10 nearest at the last layer. Prefix is the component
• linattn_z_knn_outlier: 1 minus the mean cosine to the prompt's 10 nearest prompts, averaged over layers; high means isolated. Prefix is the component
• linattn_z_hub_count: how many prompts list this one among their 10 nearest at the last layer. Prefix is the component
• mlp_knn_outlier: 1 minus the mean cosine to the prompt's 10 nearest prompts, averaged over layers; high means isolated. Prefix is the component
• mlp_hub_count: how many prompts list this one among their 10 nearest at the last layer. Prefix is the component
• q_knn_outlier: 1 minus the mean cosine to the prompt's 10 nearest prompts, averaged over layers; high means isolated. Prefix is the component
• q_hub_count: how many prompts list this one among their 10 nearest at the last layer. Prefix is the component
• up_knn_outlier: 1 minus the mean cosine to the prompt's 10 nearest prompts, averaged over layers; high means isolated. Prefix is the component
• up_hub_count: how many prompts list this one among their 10 nearest at the last layer. Prefix is the component
• v_knn_outlier: 1 minus the mean cosine to the prompt's 10 nearest prompts, averaged over layers; high means isolated. Prefix is the component
• v_hub_count: how many prompts list this one among their 10 nearest at the last layer. Prefix is the component
• attn_mahalanobis_z: Mahalanobis distance from the layer mean in the whitened top-256 PCA space, z-scored within each layer, averaged over layers; high means unusual. Prefix is the component
• attn_gate_mahalanobis_z: Mahalanobis distance from the layer mean in the whitened top-256 PCA space, z-scored within each layer, averaged over layers; high means unusual. Prefix is the component
• gate_mahalanobis_z: Mahalanobis distance from the layer mean in the whitened top-256 PCA space, z-scored within each layer, averaged over layers; high means unusual. Prefix is the component
• heads_mahalanobis_z: Mahalanobis distance from the layer mean in the whitened top-256 PCA space, z-scored within each layer, averaged over layers; high means unusual. Prefix is the component
• k_mahalanobis_z: Mahalanobis distance from the layer mean in the whitened top-256 PCA space, z-scored within each layer, averaged over layers; high means unusual. Prefix is the component
• linattn_out_mahalanobis_z: Mahalanobis distance from the layer mean in the whitened top-256 PCA space, z-scored within each layer, averaged over layers; high means unusual. Prefix is the component
• linattn_qkv_mahalanobis_z: Mahalanobis distance from the layer mean in the whitened top-256 PCA space, z-scored within each layer, averaged over layers; high means unusual. Prefix is the component
• linattn_z_mahalanobis_z: Mahalanobis distance from the layer mean in the whitened top-256 PCA space, z-scored within each layer, averaged over layers; high means unusual. Prefix is the component
• mlp_mahalanobis_z: Mahalanobis distance from the layer mean in the whitened top-256 PCA space, z-scored within each layer, averaged over layers; high means unusual. Prefix is the component
• q_mahalanobis_z: Mahalanobis distance from the layer mean in the whitened top-256 PCA space, z-scored within each layer, averaged over layers; high means unusual. Prefix is the component
• up_mahalanobis_z: Mahalanobis distance from the layer mean in the whitened top-256 PCA space, z-scored within each layer, averaged over layers; high means unusual. Prefix is the component
• v_mahalanobis_z: Mahalanobis distance from the layer mean in the whitened top-256 PCA space, z-scored within each layer, averaged over layers; high means unusual. Prefix is the component