juiceb0xc0de/bella-bartender-gemma-e4b-cgs-atlas

A brain atlas for juiceb0xc0de/bella-bartender-gemma-e4b, a single-voice fine-tune of google/gemma-4-E4B-it trained on 9,300 conversation pairs from one person, built to take the assistant reflexes out of the 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 42 layers does with them.

If you want to know where in the stack a fine-tuned Gemma 4 sorts prompts by what they are for, how much of that readout survives when whole task families are held out, why the largest direction in a mid-stack stream turns out to be a measurement artifact rather than meaning, what the 17 labels hide, and what a representation trough is and is not, this is the dataset.

This is a census run: activations only, read at two token positions and scored against 17 usage labels. It carries no weight-side circuit spectra, no logit lens, no contrast pass and no surgery, and nothing in it was measured by intervening on the model. The tables below come from two sources and say which: the atlas pipeline, and an independent pass by ARIA over the published tables and about 0.96 GB of the raw activations, run with its own protocols. The coverage section says exactly what is and is not here.

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!

What was run

• Model: juiceb0xc0de/bella-bartender-gemma-e4b at 46c7071a6c94ecdfafd83337656d186da6699b36 (revision and weights commit), architecture Gemma4ForCausalLM, base google/gemma-4-E4B-it
• Corpus: 8,965 prompts across 17 usage buckets and 175 fine-grained task families, 11 to 59 tokens, median 25, chat template applied (9 template tokens), 0 prompts truncated
• Layers probed: all 42
• Full-attention layers: 5, 11, 17, 23, 29, 35, 41 (head dim 512); sliding-window attention on the other 35 (head dim 256, window 512 tokens)
• Streams: attn, gate, heads, k, mlp, q, up, v, each read at last_input and mean_all_real. mlp is the hook mlp.down_proj:input, the 10,240-wide gated intermediate act(gate_proj x) times up_proj x, not the 2,560-wide write into the residual stream. k and v exist for layers 0 to 23 only: the model's last 18 layers share keys and values with earlier layers and have none of their own to capture
• 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
• Independent check: ARIA re-analysed the six published tables and downloaded 0.96 GB of raw activation tensors covering all 8,965 prompts at nine stream/position combinations, verified against their recorded checksums, with no changes to any repository. Its probes use 128 principal components and either one stratified 75/25 split or five folds grouped by task family; they are additional protocols, not reproductions of the atlas's 256-component, five-fold evaluation, and every number from them is marked ARIA below
• Not run: OV-circuit SVD, logit lens, contrast datasets, Sub-Zero surgery, code analysis

Architecture notes

Property	Value
Architecture	Gemma4ForCausalLM
Base model	google/gemma-4-E4B-it
Fine-tune corpus	9,300 conversation pairs, one voice
Hidden size	2,560
Intermediate size	10,240
Layers	42
Query heads	8
KV heads	2
Head dimension	256 on sliding layers, 512 on full-attention layers
Full-attention cadence	every sixth layer, starting at 5
Sliding window	512 tokens
KV-shared layers	18 (layers 24 to 41)
Vocabulary	262,144
Stream widths	attn 2,560; gate, up, mlp 10,240; heads and q 8 x 512; k and v 2 x 512

Corpus

bucket	prompts
business	590
design	590
humor	590
niche	590
introspection	589
roleplay	589
community	588
core_technical	550
ml_ai	550
research	527
creative_writing	515
writing	514
brainstorming	499
planning	472
data	467
learning	455
tool_use	290

8,965 prompts, 11 to 59 tokens, median 25, the same corpus as every atlas under this account. The 17 buckets are the labels every per-label number uses; underneath them the corpus file carries 175 fine-grained task families in fine_bucket, which matter for finding 1. The largest bucket is 6.58% of the corpus, which is the chance level for every probe below.

What the tables contain

Table	Rows	What it gives you
features	1,627,136	one row per (layer, component, neuron): ANOVA F, q, eta2, top label, AUROC, association, taxonomy
layers	600	one row per (layer, component, position): sanity, geometry, probe, taxonomy counts, CKA with the other streams, neighbours, trajectories
per_head	768	per-head output-norm F, q and top label for heads, q, k and v
coactivation	15,000	the 50 strongest correlated pairs among the 256 highest-F neurons, per stream per layer
dictionary_features	1,024	128 learned atoms per stream at its most selective layer: frequency, activation, top label, AUROC
prompts	8,965	the corpus with labels, task families, token counts, and per-stream outlier, hub and Mahalanobis scores

atlas/atlas.sqlite holds the same six tables.

from datasets import load_dataset

repo = "juiceb0xc0de/bella-bartender-gemma-e4b-cgs-atlas"
features = load_dataset(repo, "features", split="train")
layers = load_dataset(repo, "layers", split="train")

To find the neurons that tell the labels apart, sort features by neg_log10_q and threshold on q (significant below 0.01); taxonomy says whether a neuron is selective for one label, mixed across several, or broad.

Key findings

1. Usage category is readable from layer 13 to 17, and the headline depends on the split

The atlas's held-out probe, a multinomial logistic regression on the top 256 principal components over 5 stratified folds, chance 0.066:

stream	layer	accuracy	shuffled-label control	selectivity
mlp	13	0.839	0.063	0.777
gate	13	0.827	0.059	0.768
heads	17	0.804	0.060	0.744

The layer where each stream's probe is most selective, which is also where its dictionary was learned:

attn	gate	mlp	up	heads	k	q	v
13	13	13	13	17	17	17	17

The MLP-side streams peak at 13 and the attention-side streams at 17, and a reader of the labels out of either gets four prompts in five right against a control that gets one in sixteen. That is the number as the atlas measures it, on a random split of prompts. The corpus, though, is built from 175 task families, and a random split puts prompts from the same family on both sides of the fold. ARIA asked the harder question, whether the readout holds up when entire families are held out:

representation (ARIA)	random prompt split	task families held out
word and bigram TF-IDF	76.0%	36.7%
layer-13 gate, 128 PCs	77.7%	50.3%
layer-13 mlp intermediate, 128 PCs	79.1%	51.5%
prompt length alone	7.9%
majority class	6.58%

Holding families out costs the activation probes about 27 points and the lexical baseline about 39, and on families the probe has never seen the activations still lead the words by 14 to 15 points. So the 84% is not evidence of 84% general-purpose intent recognition, and it is not a length classifier either, since length alone sits a point above chance. What the atlas supports is category information with measurable generalisation across task families: a narrower claim than the headline, and a more useful one, because it comes with a number for how much of the readout is the taxonomy and how much is the model.

2. Two bands, with a trough at layer 23

Share of a prompt's 10 nearest neighbours (cosine, top-256 PCA space) that carry its label:

stream	first peak	layer 23	second peak
attn	0.613 (layer 13)	0.292	0.616 (layer 30)
heads	0.668 (layer 13)	0.337	0.573 (layer 30)
gate	0.608 (layer 13)	0.272	0.526 (layer 32)
up	0.629 (layer 13)	0.306	0.539 (layer 32)
mlp	0.535 (layer 13)	0.289	0.467 (layer 31)
k	0.574 (layer 14)	0.420	not captured past 23
q	0.556 (layer 17)	0.363	not captured past 23
v	0.574 (layer 14)	0.407	not captured past 23

Every stream sorts prompts by label best in the low to mid teens, loses most of it by layer 23, and recovers only part of it in a second band around layers 30 to 32; the attention output is the one stream whose second band matches its first. The obvious worry is that the trough is a length effect, since prompt length is the largest direction in these streams (finding 6). ARIA probed the attention output on either side of the trough from the raw activations, before and after regressing out length terms:

attention output, mean position (ARIA, 128 PCs)	layer 22	layer 23	layer 24
held-out accuracy	71.8%	55.5%	73.3%
after length adjustment	71.5%	55.9%	72.8%

The trough is not the length terms. But "the model loses category information at layer 23" would still be too strong. These are component outputs, not the residual stream, so the information can be sitting elsewhere in the block, and at the final input position layer 23's attention output still supports 64.0% held-out accuracy in ARIA's corresponding probe, despite being only four directions wide there (finding 7). Low effective dimension is not the same thing as no useful information.

3. In the back half, every full-attention layer is a dip

Label agreement at the seven full-attention layers, and whether it sits below the neighbouring layers:

full-attention layer	attn	heads	local minimum
5	0.429	0.465	no
11	0.534	0.527	no
17	0.466	0.600	no
23	0.292	0.337	yes
29	0.297	0.351	yes
35	0.258	0.329	yes
41	0.321	0.402	yes

The front three full-attention layers sit at or above their neighbours. The back four are each a hole in the label structure, for the attention output and for the heads alike. One architectural fact rules out the easy explanation: the sliding window is 512 tokens and the longest prompt here is 59, so on this corpus the sliding layers and the full-attention layers see the same context, all of it. Whatever the back-half full-attention layers are doing differently, it is not reaching tokens the sliding layers cannot. Different weights, the 512 head dimension against 256, the positional scheme and the roles the layers settled into during training are all live candidates, and telling them apart needs measurements this atlas does not take.

4. Adjacent-layer redundancy sits in the MLP-side streams and the keys, and some of it is the shared direction

Adjacent layers with linear CKA at or above 0.95, the breaks between them, and the natural layer blocks:

stream	redundant adjacent runs	breaks	natural blocks
gate	8 to 10, 11 to 17, 18 to 22, 24 to 29, 30 to 40	10/11, 17/18, 22/23, 23/24, 29/30, 40/41	0; 1 to 5; 6 to 7; 8 to 10; 11 to 12; 13 to 17; 18 to 20; 21 to 23; 24 to 29; 30 to 40; 41
up	8 to 10, 11 to 12, 13 to 22, 24 to 29, 30 to 41	10/11, 12/13, 22/23, 23/24, 29/30	0; 1 to 2; 3 to 5; 6 to 7; 8 to 10; 11 to 12; 13 to 17; 18 to 22; 23 to 29; 30 to 40; 41
mlp	2 to 17, 18 to 22, 27 to 29	17/18, 22/23, none after 29	0; 1 to 17; 18 to 23; 24 to 29; 30 to 31; 32 to 35; 36 to 40; 41
k (layers 0 to 23)	5 to 7, 8 to 11, 12 to 13, 14 to 15, 20 to 21	7/8, 11/12, 13/14, 15/16	0; 1 to 13; 14 to 15; 16 to 17; 18 to 21; 22 to 23
attn	none		0 to 1; 2 to 5; 6 to 17; 18 to 23; 24 to 29; 30 to 32; 33 to 35; 36 to 39; 40 to 41
heads	none		0 to 2; 3 to 5; 6 to 7; 8 to 12; 13 to 23; 24 to 29; 30 to 31; 32 to 36; 37 to 40; 41
q	none		0 to 1; 2 to 3; 4 to 7; 8 to 10; 11 to 12; 13 to 17; 18 to 23; 24 to 29; 30 to 32; 33 to 37; 38 to 41
v (layers 0 to 23)	none		0 to 1; 2 to 6; 7 to 11; 12 to 13; 14 to 15; 16 to 18; 19 to 22; 23

As measured, adjacent-layer redundancy belongs to gate, up, mlp and the keys; the attention output, the heads, the queries and the values never reach 0.95 between neighbours anywhere in the stack. The breaks in the gate and up runs fall at 10/11, 17/18, 22/23/24 and 29/30, each beside a full-attention layer (11, 17, 23, 29), while the run from 30 to 40 passes over full-attention layer 35 without a break. What the threshold means needs care, though. ARIA reproduced one pair from the raw activations and then removed the length-related terms from both sides:

layer 12 against layer 13, mlp (ARIA, 2,048 prompts)	linear CKA
raw representations	0.9958
after length adjustment (L, 1/L, L squared regressed out)	0.8879

Still a lot of shared geometry, but no longer over the atlas's redundancy line. The honest reading of the runs in the table is "highly similar measured geometry, partly carried by a shared length-related contribution", not "these layers do the same computation" and not "these layers can be removed". CKA never established interchangeability, and this adjustment was checked on one adjacent pair, not on every run reported.

5. At the final input token, layer 0 is one direction

Streams with a participation ratio below 3, where a single direction holds the whole stream:

layer	stream	position	participation ratio
0	attn	last_input	1.67
0	gate	last_input	2.507
0	heads	last_input	1.751
0	k	last_input	1.435
0	mlp	last_input	1.949
0	mlp	mean_all_real	2.61
0	q	last_input	1.0
0	up	last_input	2.315
0	v	last_input	1.471
1	k	last_input	2.814
1	q	last_input	2.34
1	v	last_input	2.254
10	mlp	mean_all_real	2.423
11	mlp	mean_all_real	2.905
35	attn	mean_all_real	2.097

Twelve of the fifteen rows are the final input position, and that is the template, not the model. With the chat template applied, the last input token is the same newline (token 107, the end of the generation prompt) for every one of the 8,965 prompts, so at layer 0, where a stream is a function of the token and its position and nothing else, every stream at that position is close to constant across prompts. The queries come out as exactly one direction. For the same reason this does not carry to the mean position: ARIA's read of the raw activations puts layer 0 at the mean position at 18.79 for the attention output, 20.94 for gate and 34.59 for heads. The one mean-position stream that is one direction at layer 0 is the mlp intermediate, at 2.61, and it is one direction again at layers 10 and 11. Finding 6 says what that direction is. For every row in the table, the geometry, trajectory and neighbour numbers describe the dominant direction and nothing else; the probe is unaffected. The attention output at layer 35, the one late entry, is also the stream that is most hub-dominated at that layer (finding 11).

6. The largest direction in the layer-13 mlp stream is a pooling artifact

Correlation of the first three principal components with prompt length, as the atlas reports it:

stream (layer)	PC1	PC2	PC3
mlp (13)	0.929	0.064	0.000
gate (13)	0.181	0.315	0.309
heads (17)	0.527	0.060	0.232

The atlas also puts one neuron at the top of the layer-13 mlp outlier list at the mean position:

neuron	mean absolute activation	median z-score
9239	16.5	551.1

ARIA went to the raw activations and made the first row precise:

layer-13 mlp intermediate, mean position (ARIA)	result
variance explained by PC1 (training split)	33.8%
correlation of PC1 with sequence length, absolute	0.929
correlation of PC1 with inverse sequence length	0.9983
correlation of neuron 9239 with inverse sequence length	0.9989
share of PC1's squared loading on neuron 9239	71.7%
neuron 9239 activation times sequence length	about 394.9
coefficient of variation of that product	1.08%

The first axis of the stream is not prompt length. It is one neuron whose activation times the number of tokens is a constant to within one percent, which is what a fixed per-sequence contribution looks like after it has been divided by the sequence length in mean pooling: the mean over L tokens of something that is only there once comes out as c over L. A shared template or special-token contribution is the plausible source, and the pooled captures cannot say which token it is, so that stays a hypothesis. What they can say is what happens to the label structure when the direction is taken out:

layer-13 mlp, held-out category-neighbour agreement (ARIA)	result
original	49.0%
with PC1 removed	57.0%
with length-related terms regressed out	57.3%

Removing the largest geometric feature of the stream improves the semantic structure. That is the finding the atlas can make that a plain probe cannot: it can expose when the dominant direction in a stream is a property of the measurement rather than of the model, and the rows in finding 5 where the mean-position mlp intermediate is one direction, at layers 0, 10 and 11, are the natural places to look for the same mechanism. The chart above says it is not three layers but a region: at the mean position the mlp intermediate stays under a participation ratio of 14 at every layer from 2 through 17, while the other streams sit between 20 and 100, and it opens up to 78 only at layer 24. The attention output does the opposite, collapsing from layer 35 to the end, which is finding 11.

7. Full attention collapses the attention output at the final input token

Participation ratio of the attention output at last_input, at each full-attention layer and its neighbours:

full-attention layer	layer before	at the layer	layer after
5	45.6	12.6	28.5
11	37.2	39.8	49.5
17	10.0	3.15	23.1
23	9.68	4.33	49.0
29	59.1	22.4	65.2
35	36.9	8.41	9.45
41	27.0	20.1

Across the whole stack the same number runs from 1.67 at layer 0 to a peak of 76.8 at layer 26. Five of the seven full-attention layers pull it down against both neighbours, and at 17 and 23 the attention output at the last token is three or four directions wide. Narrow is not empty: at layer 23, where the participation ratio is 4.33, ARIA's probe at this position still reads the usage category at 64.0% held out. A handful of directions is enough to carry the label, which is worth remembering before reading any participation ratio in this atlas as a measure of content.

8. Layer 5 is the loudest attention layer

Median attention-output norm at the mean position, layers 0 to 7:

layer	0	1	2	3	4	5	6	7
median norm	48.2	131.9	84.2	48.0	89.2	270.9	166.5	64.6

The first full-attention layer writes three times as much as the layer before it, and the surge carries into layer 6. No layer from 8 to 39 is above 75.

9. Significance is cheap, selectivity is not, and three neurons are worth a closer look

Neuron taxonomy at layer 13, the most selective MLP-side layer:

stream	width	selective	mixed	broad	inert
attn	2,560	132	6	2,422	0
gate	10,240	1,085	312	8,842	1
heads	2,048	199	19	1,830	0
k	512	19	1	492	0
mlp	10,240	408	61	9,770	1
q	2,048	95	13	1,940	0
up	10,240	1,042	222	8,976	0
v	512	38	2	472	0

Broad means q below 0.01 but no single label with an AUROC margin of 0.25: with 8,965 prompts nearly every neuron responds to the labels, and few respond to one. The selective counts keep rising one layer further, to 1,689 for gate at layer 14. At the other end, layer 0's attention output has 2,538 of 2,560 neurons broad, and the mlp intermediate at layer 5 has 1,763 neurons that do not respond to the labels at all. Among the selective gate neurons, ARIA checked whether the strongest associations are secretly length:

gate neuron, layer 13 (ARIA)	associated label	full-corpus AUROC	correlation with length
2264	creative_writing	0.925	0.013
8246	ml_ai	0.935	0.036
5015	core_technical	0.911	0.035

They are not, and the associations hold after ARIA's length adjustment. The examples say something about the labels as well: neuron 2264 fires on creative-writing prompts and on the surrealist-writing prompts the corpus files under humor, which fits the cosine of 0.792 the atlas measures between the creative_writing and humor directions. These are overlapping motifs, not cleanly separated capability boxes, and the right name for the three neurons is candidate task-associated features. Whether they do anything for the model's behaviour needs an ablation, a patch or a steering run with an output-level measurement, none of which this atlas contains.

10. The dictionaries are organised, just not by the 17 labels

128 atoms per stream, learned at the layer where that stream's probe is most selective:

stream	layer	dead	dense	label-aligned
attn	13	0	0	0
gate	13	0	0	1
mlp	13	0	0	2
up	13	0	0	1
heads	17	0	0	1
k	17	0	0	0
q	17	0	0	0
v	17	0	0	0

Of 1,024 atoms, none is dead, none is dense, and 5 clear the coarse-label alignment threshold, all of them for tool_use, the smallest and most distinct bucket. Read on its own that says the dictionaries ignore the corpus's categories. ARIA read the stored top-activation examples instead, and the picture inverts:

top-example coherence, layer 13 (ARIA)	gate	mlp
mean share of an atom's top ten examples from one majority label	73.5%	79.0%
atoms with at least 8 of 10 top examples from one task family	55 of 128	63 of 128

atom (ARIA)	what its top examples are	coarse-label AUROC
gate atom 2	all ten are tagline generation	0.573
mlp atom 2	all ten are villain monologues	0.561
mlp atom 37	all ten are workflow automation design
gate atom 23	agent tool selection and tool-use prompting, across niche and tool_use

Half the gate dictionary and more than half the mlp dictionary are organised around task families finer than anything the 17 labels can express, and a one-versus-rest AUROC against a coarse label cannot see it. Two cautions go with that. These dictionaries were learned in standardized PCA space, not directly on neurons, and their reconstruction R squared is about 0.353 for gate and 0.412 for mlp, so they are partial summaries of the streams. And top-example coherence inside one corpus is not monosemanticity. It is enough to say that the useful unit here is the atom with its prompts, not the atom with its AUROC.

11. Late attention output is hub-dominated

Hubness skew of the attention output at the mean position, the three highest layers:

layer	hubness skew
31	6.13
35	7.26
41	6.77

A few prompts are everyone's neighbour in the late attention output, where the gate stream's hubness skew stays between 1.04 and 1.88 at every one of the 42 layers. Layer 35 is also the one late stream in finding 5 that collapses to a single direction at the mean position, at a participation ratio of 2.097: the same layer, the same position, one direction and a handful of hubs.

How to read the numbers

• Two sources. Every table is the atlas pipeline's unless it says ARIA. ARIA's figures come from its own protocols on the raw captures, 128 principal components with either one stratified 75/25 split or five folds grouped by task family, and from the published tables; they are additional measurements, not reproductions of the atlas's 256-component, five-fold probe.
• mlp is the gated intermediate. The stream is mlp.down_proj:input, act(gate_proj x) times up_proj x, 10,240 wide. It is not the MLP's 2,560-wide contribution to the residual stream, and no residual-space quantity is in this atlas.
• AUROC runs both ways. A top_auroc below 0.5 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.
• Labels overlap. They are usage categories, not mutually exclusive 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, and the per-label directions pull toward each other.
• 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 comes from a log-space tail of the F distribution and stays finite for every finite F; it is infinite only when every label's rows share one value.
• Probe accuracy is 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 and the same per-fold PCA.
• Positions. last_input: the activation at the final input position (token_pos = seq_len - 1), which with the chat template is the last token of the generation prompt (a newline, token 107); the model has produced nothing at this point. mean_all_real: the mean over every real (unpadded) token, template and special tokens included, which is what lets a fixed per-sequence contribution show up as an inverse-length direction (finding 6). Per-neuron and per-prompt numbers use mean_all_real. Older atlases call these last and mean.
• Every number is descriptive. Nothing here was measured by intervening on the model.
• Non-finite JSON numbers are null; dictionary values carry a sibling <key>_status of positive_infinity, negative_infinity or nan. NumPy and Parquet retain the values.

Coverage

Captured: the eight streams above at both positions, for every layer the model has them; k and v stop at layer 23 because layers 24 to 41 reuse earlier layers' keys and values rather than computing their own. Not captured, so not in this atlas: residual states, next-token predictions, spike statistics, attention-head summaries, contrast datasets, weight inputs, Sub-Zero passes. Because there is no residual stream there is no cos_to_end and no settling layer; a component output has no end state to converge to. The raw activation captures the pipeline and ARIA both read from sit in the atlas bucket under jobs/bella-bartender-gemma-e4b-census-261019-2f1, about 58 GB in all.

Files

Paths are relative to atlas/.

• data/<table>.parquet and atlas.sqlite: the six tables
• manifest.json: model shape, components per layer and corpus labels; _manifest.json: table schemas with row counts; model_meta.json: model, revision, capture version, fingerprint and per-layer attention types; capture_analysis.json: which analyses ran and what was not captured
• layers/<l>/meta.json: the layer's attention type, heads and streams
• layers/<l>/components/<comp>/summary.json: every analysis's numbers for one stream, both positions
• layers/<l>/components/<comp>/separation_scores.npy: one-way ANOVA F per neuron against the labels
• layers/<l>/components/<comp>/directions.npy: one difference-of-means direction per label, unit length, float16
• layers/<l>/components/<comp>/bucket_metrics.json: per label, its rows, mean norm, selective neurons and top neurons by z; cosines between label directions
• layers/<l>/components/<comp>/taxonomy.json: taxonomy counts and the strongest examples of each class
• layers/<l>/components/<comp>/coactivation.json: the 50 strongest correlated pairs among the 256 highest-F neurons
• layers/<l>/components/<comp>/dictionary.json: the learned dictionary, at the one layer per component where the probe is most selective
• layers/<l>/components/<comp>/census_heatmap.png: label mean z-score of the 64 highest-F neurons
• layers/<l>/component_comparison.json: the components of one layer side by side, with CKA between them
• cross_layer/norms.json: median row norm per layer
• cross_layer/geometry.json: participation ratio, TwoNN, anisotropy and k90 per layer, and the dominated streams
• cross_layer/trajectories.json: consecutive-layer cosine, update size, cos_to_end and the end layer used
• cross_layer/cka_by_layer.json, cka.png: layer by layer linear CKA per stream, with natural layer blocks and redundant adjacent pairs (CKA at or above 0.95)
• cross_layer/neighbors.json: kNN label agreement, k-means NMI, neighbour overlap with the previous layer, hubness
• cross_layer/fstat_top_by_layer.json: the 20 highest-F neurons per layer and component
• cross_layer/taxonomy_by_layer.json: taxonomy counts per layer and component
• cross_layer/direction_stability.json: cosine of each label's direction with the layer before
• cross_layer/dictionary_layers.json: which layer each component's dictionary was learned at

Columns

Every table except prompts carries layer (layer index, 0 is the first block's output) and component (the stream: attn, gate, heads, k, mlp, q, up, v).

features

• 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. 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

• position: which token position the stream was read at: mean_all_real or last_input
• sanity_rows, sanity_features: prompts in the stream and its width
• sanity_nan, sanity_inf: NaN and infinite values found
• sanity_norm_p05, sanity_norm_p50, sanity_norm_p95: 5th percentile, median and 95th percentile of row L2 norm
• sanity_linf_p50: median over rows of the row's largest |value|
• sanity_row_skew_p50, sanity_row_kurtosis_p50: skew and excess kurtosis of a row's values across its features, median row
• sanity_frac_abs_lt_1e-3, sanity_frac_abs_lt_1e-2: share of all values with |x| below 0.001 and 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, geometry_k90, geometry_k95, geometry_k99: principal components needed to reach 50%, 90%, 95% and 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 neighbour numbers describe that direction only
• geometry_pca_explained: share of the variance in the top 256 principal components
• trajectories_mahalanobis_p50, trajectories_mahalanobis_p99: median and 99th percentile Mahalanobis distance of a prompt from the layer's mean in the whitened top-256 PCA space
• trajectories_cos_to_mean_p50: median cosine between a row and the layer's mean row
• trajectories_cos_prev_mean, trajectories_cos_prev_p05: mean and 5th percentile cosine between a prompt's state here and one layer earlier
• 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
• layers_xcka_cka_with_<stream>.mean_all_real, one column per stream: 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 neighbour
• neighbors_knn_cos_mean: mean cosine to the 10 nearest prompts
• neighbors_knn_jaccard_prev: Jaccard overlap between a prompt's 10 nearest here and one layer earlier, averaged over prompts
• neighbors_duplicates: prompt pairs with cosine at or above 0.98 in the top-256 PCA space
• neighbors_kmeans_ari, neighbors_kmeans_nmi: adjusted Rand index and normalized mutual information of k-means (one cluster per label) against the labels
• neighbors_silhouette_labels: silhouette of the labels under cosine, on a 2000-row subsample
• probes_probe_accuracy, probes_probe_accuracy_std: held-out accuracy of a multinomial logistic regression on the top-256 principal components, 5 stratified folds, and its standard deviation across 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, probes_max_F: median and largest F over the stream's neurons
• probes_frac_significant: share of neurons with q below 0.01
• probes_eta2_mean: mean eta2 over neurons
• probes_taxonomy_<class>, one column per class (dead, inert, selective, mixed, broad): count of neurons in that class
• 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

per_head

• 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, dim_F_max: mean and largest F over the head's dimensions
• dim_frac_significant: share of the head's dimensions with q below 0.01

coactivation

• 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

• 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

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, subcategory, fine_bucket: top-level, second-level and finest grouping from the corpus file, when the corpus has them
• 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
• <comp>_knn_outlier, one column per stream: 1 minus the mean cosine to the prompt's 10 nearest prompts, averaged over layers; high means isolated
• <comp>_hub_count, one column per stream: how many prompts list this one among their 10 nearest at the last layer
• <comp>_mahalanobis_z, one column per stream: Mahalanobis distance from the layer mean in the whitened top-256 PCA space, z-scored within each layer, averaged over layers; high means unusual

Provenance

• capture format census version 2, fingerprint sha256:c26b0de433d517f6f9d61e43e33a109c4770142f48d9aa0c869fae2cc77ee7c8
• capture code 57da548de6c50db1452148985e4b97eefbaeab83, analysis code f3fe1ba0fd604f745f80ec878481fbb2d0a72666
• 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
• job bella-bartender-gemma-e4b-census-261019-2f1
• independent check: ARIA, on the six published tables and 0.96 GB of raw activation tensors (all 8,965 prompts, nine stream/position combinations) verified against their recorded checksums, with its own protocols and no changes to any repository