Datasets:
layer int64 0 23 | component stringclasses 12
values | a int64 0 6.1k | b int64 31 6.14k | r float64 -0.97 0.98 |
|---|---|---|---|---|
0 | gate | 4,194 | 5,928 | 0.8086 |
0 | gate | 1,361 | 6,139 | 0.7982 |
0 | gate | 1,361 | 5,875 | 0.7937 |
0 | gate | 2,256 | 5,438 | 0.7845 |
0 | gate | 16 | 3,508 | 0.783 |
0 | gate | 3,193 | 3,930 | 0.781 |
0 | gate | 3,624 | 4,335 | 0.7807 |
0 | gate | 3,930 | 4,194 | 0.7806 |
0 | gate | 1,100 | 5,499 | 0.7726 |
0 | gate | 1,361 | 4,648 | 0.7691 |
0 | gate | 3,624 | 4,406 | 0.7685 |
0 | gate | 4,320 | 5,499 | 0.7659 |
0 | gate | 2,256 | 5,507 | 0.7624 |
0 | gate | 4,648 | 5,875 | 0.7611 |
0 | gate | 2,867 | 3,674 | -0.7587 |
0 | gate | 653 | 4,648 | 0.7582 |
0 | gate | 3,853 | 4,320 | 0.7582 |
0 | gate | 4,194 | 5,499 | 0.7573 |
0 | gate | 2,256 | 5,965 | 0.7569 |
0 | gate | 4,648 | 5,289 | 0.7563 |
0 | gate | 1,782 | 4,777 | 0.7557 |
0 | gate | 5,875 | 6,139 | 0.7546 |
0 | gate | 4,648 | 6,139 | 0.7532 |
0 | gate | 143 | 3,171 | 0.7517 |
0 | gate | 143 | 4,194 | 0.7505 |
0 | gate | 3,065 | 4,335 | 0.7485 |
0 | gate | 5,304 | 5,507 | 0.7476 |
0 | gate | 5,966 | 6,139 | 0.7473 |
0 | gate | 4,406 | 5,438 | 0.7469 |
0 | gate | 4,219 | 6,076 | 0.7463 |
0 | gate | 2,888 | 3,624 | 0.7439 |
0 | gate | 2,867 | 4,883 | 0.7437 |
0 | gate | 4,777 | 4,880 | 0.7426 |
0 | gate | 2,355 | 3,207 | 0.7424 |
0 | gate | 2,867 | 4,197 | 0.7411 |
0 | gate | 5,438 | 5,965 | 0.7409 |
0 | gate | 4,219 | 4,648 | 0.7408 |
0 | gate | 16 | 3,489 | 0.7402 |
0 | gate | 3,171 | 4,194 | 0.7398 |
0 | gate | 1,788 | 3,021 | 0.7376 |
0 | gate | 1,100 | 4,194 | 0.7373 |
0 | gate | 1,825 | 3,065 | 0.7364 |
0 | gate | 5,417 | 5,518 | 0.7361 |
0 | gate | 653 | 5,875 | 0.7353 |
0 | gate | 1,361 | 3,087 | 0.734 |
0 | gate | 4,648 | 5,242 | 0.7333 |
0 | gate | 3,087 | 4,648 | 0.7324 |
0 | gate | 5,304 | 5,965 | 0.7317 |
0 | gate | 3,193 | 4,194 | 0.7315 |
0 | gate | 777 | 3,171 | 0.7306 |
0 | linattn_out | 367 | 1,492 | 0.6435 |
0 | linattn_out | 127 | 536 | 0.6153 |
0 | linattn_out | 574 | 783 | -0.517 |
0 | linattn_out | 367 | 1,813 | 0.4966 |
0 | linattn_out | 435 | 1,492 | 0.4817 |
0 | linattn_out | 1,010 | 1,747 | 0.4771 |
0 | linattn_out | 367 | 2,022 | 0.4669 |
0 | linattn_out | 424 | 536 | 0.4477 |
0 | linattn_out | 745 | 1,492 | 0.4467 |
0 | linattn_out | 367 | 1,472 | 0.4466 |
0 | linattn_out | 1,019 | 1,496 | 0.444 |
0 | linattn_out | 307 | 1,747 | 0.4353 |
0 | linattn_out | 127 | 424 | 0.4223 |
0 | linattn_out | 752 | 1,492 | 0.4204 |
0 | linattn_out | 790 | 1,727 | -0.4187 |
0 | linattn_out | 97 | 1,492 | 0.4167 |
0 | linattn_out | 367 | 854 | 0.4145 |
0 | linattn_out | 929 | 1,492 | -0.4117 |
0 | linattn_out | 574 | 1,951 | -0.409 |
0 | linattn_out | 1,010 | 1,727 | 0.4073 |
0 | linattn_out | 367 | 945 | -0.407 |
0 | linattn_out | 367 | 1,102 | -0.405 |
0 | linattn_out | 783 | 1,951 | 0.4029 |
0 | linattn_out | 929 | 1,618 | -0.4024 |
0 | linattn_out | 574 | 929 | -0.3998 |
0 | linattn_out | 790 | 1,591 | 0.3991 |
0 | linattn_out | 1,733 | 1,934 | -0.3983 |
0 | linattn_out | 424 | 1,747 | -0.3975 |
0 | linattn_out | 726 | 1,492 | 0.3973 |
0 | linattn_out | 367 | 1,067 | -0.3924 |
0 | linattn_out | 389 | 1,951 | -0.3921 |
0 | linattn_out | 367 | 1,386 | -0.3896 |
0 | linattn_out | 1,492 | 1,934 | -0.3895 |
0 | linattn_out | 745 | 1,934 | -0.3894 |
0 | linattn_out | 139 | 1,010 | 0.3893 |
0 | linattn_out | 367 | 478 | 0.3888 |
0 | linattn_out | 127 | 815 | -0.3874 |
0 | linattn_out | 929 | 1,496 | -0.387 |
0 | linattn_out | 1,951 | 1,969 | -0.387 |
0 | linattn_out | 233 | 745 | -0.3863 |
0 | linattn_out | 367 | 1,676 | -0.3861 |
0 | linattn_out | 367 | 934 | -0.3847 |
0 | linattn_out | 307 | 1,010 | 0.3846 |
0 | linattn_out | 367 | 536 | -0.3841 |
0 | linattn_out | 1,618 | 1,951 | -0.3833 |
0 | linattn_out | 1,025 | 1,934 | 0.383 |
0 | linattn_out | 536 | 1,120 | -0.3807 |
0 | linattn_out | 1,719 | 1,747 | 0.3798 |
0 | linattn_out | 777 | 1,492 | 0.3779 |
0 | linattn_out | 391 | 1,010 | 0.3778 |
- Selective MLP entries fall 73.71% and probe accuracy declines at every layer
- Every label loses selective entries; humor gains 7.21 points of share
- Coordinate rankings and selected partnerships both change substantially
- MLP PR expands at 15 of 16 early and middle layers, then contracts at all 8 late layers
- 18 linear-attention blocks and 6 full-attention blocks are mapped
- Verdict: decision training accompanies broad backbone reorganization
- Limits
Joey-1: a decision-model fine-tune through a general prompt census
I ran the joey-1 and parent Qwen3.5-2B models through the Cartographics HF space (a space that generates the model atlases referenced throughout this article, on demand). Base Labs describes joey-1 as a full language-model fine-tune on 6.21M text and image decisions, with a frozen vision encoder and an objective that includes probability calibration. This census asks how that decision training reshapes general category organization in the text backbone.
Selective MLP coordinates fall from 5,956 to 1,566, a 73.71% reduction, while mean category-probe accuracy moves from 76.04% to 68.89%. [data] I tested whether this general census showed increased single-label selectivity for tool_use or data; neither category showed an increase under this threshold. [data] Mean probe accuracy is 7.15 points lower, and every layer declines, with the largest gap at layer 23: 13.14 points. [data]
Scope. This census measures text-input activations for 8,965 prompts across 17 general usage labels. It has no calibrated-decision or abstention label and does not measure images, option probabilities, or decision accuracy. The public capture uses a generic
AutoModelForCausalLMpath rather than Joey's documented OneJev decision-serving configuration. Unless stated otherwise, results use the MLP intermediate stream atmean_all_real, averaging all real input tokens, including template and special tokens. The documented usage-label probe is a five-fold, 256-PC multinomial classifier with training-fold PCA and scaling; the majority-class baseline is 6.58%.
Evidence tags distinguish analytical claims: [data] is a direct measurement or calculation from the pinned atlas databases or capture metadata, [inferred] is my interpretation under stated assumptions, [data, inferred] is both, and [guess] is a mechanism I haven't tested.
Selective MLP entries fall 73.71% and probe accuracy declines at every layer
Selective entries fall from 4.04% to 1.06% of the 147,456 MLP layer-coordinate entries, while mean probe accuracy changes by −7.15 percentage points. [data] The atlas calls a coordinate selective when its category ANOVA has Benjamini-Hochberg-adjusted q < 0.01 and exactly one label reaches |AUROC − 0.5| ≥ 0.25; negative associations count too. Mixed means at least two labels reach that effect threshold.
| Layers, zero-based | Selective entries: parent to fine-tune | Probe accuracy: parent to fine-tune | Shuffled-label control: parent to fine-tune |
|---|---|---|---|
| 0 to 7 | 870 to 262 | 74.70% to 71.30% | 6.17% to 6.27% |
| 8 to 15 | 1,174 to 161 | 75.05% to 66.86% | 6.11% to 6.27% |
| 16 to 23 | 3,912 to 1,143 | 78.37% to 68.50% | 6.15% to 6.10% |
| All 24 layers | 5,956 to 1,566 | 76.04% to 68.89% | 6.14% to 6.21% |
The middle third loses 86.29% of its selective entries, and probe accuracy is lower at 24 of 24 layers. [data] Paired per-prompt probe errors are absent from the exports, so these accuracy gaps cannot be subjected to a paired significance test here.
Joey's highest MLP probe accuracy is 75.44% at layer 5, with a 6.25% shuffled control and 69.19-point selectivity. The parent's MLP peak is 79.33% at layer 20, with a 6.03% control and 73.30-point selectivity. [data] The fine-tune's strongest mean-pooled MLP readout occurs earlier in the stack. [data, inferred]
Category-mean z-scores for selected high-F coordinates at layer 5.
Every label loses selective entries; humor gains 7.21 points of share
tool_use falls from 869 to 209, data from 74 to 17, and core_technical from 530 to 202 selective MLP entries. [data] Brainstorming and niche have the largest proportional declines, 15 to 0 and 6 to 0. Humor has the smallest, 525 to 251 (−52.19%). [data] Zero selective entries means that no coordinate cleared this export's single-label threshold for that label, not that the label has no measurable response.
| Strongest label association | Parent count | Joey count | Parent share | Joey share |
|---|---|---|---|---|
tool_use |
869 | 209 | 14.59% | 13.35% |
creative_writing |
842 | 166 | 14.14% | 10.60% |
roleplay |
733 | 182 | 12.31% | 11.62% |
ml_ai |
730 | 128 | 12.26% | 8.17% |
introspection |
694 | 198 | 11.65% | 12.64% |
core_technical |
530 | 202 | 8.90% | 12.90% |
humor |
525 | 251 | 8.81% | 16.03% |
data |
74 | 17 | 1.24% | 1.09% |
Humor gains 7.21 percentage points of share while losing 52.19% of its entries because its count decline is the smallest in the set. Core_technical gains 4.00 points of share despite a 61.89% count decline. Ml_ai has the largest share loss, 4.08 points, after an 82.47% count decline. [data] Positive and negative selective associations remain nearly balanced: 2,999 / 2,957 in the parent and 763 / 803 in Joey. [data]
Across all captured components, 13,503 selective and 2,573 mixed entries make up 16,076 strong-association entries, down from 51,730 in the parent; 97.61% of Joey's 681,984 feature entries are broad. [data] Projected streams overlap computationally, so these are layer/component entries rather than a count of distinct physical units.
Coordinate rankings and selected partnerships both change substantially
Corresponding MLP-coordinate F ranks have median within-layer Spearman correlation 0.496. Band medians move from 0.625 in layers 0 to 7 to 0.467 in layers 8 to 15 and 0.427 in layers 16 to 23; layer 23 is lowest at 0.347. [data] Median overlap of the top-256-F coordinate sets is 33.79%, with a minimum of 26.56% at layer 11. [data]
Only 137 of 1,200 exported MLP coactivation pairs recur: 11.42% retention, with pooled Jaccard overlap 0.061. [data] 402 parent pairs retain both endpoints in Joey's top-256 sets, so endpoint survival is 33.50% while pair recurrence is 11.42%. [data] The selected partnership lists change more often than the high-F endpoint sets. [data, inferred] Full pair-correlation matrices would test whether the missing edges left the strongest list or reflect larger correlation changes; the export alone cannot separate those cases.
In the full-attention heads stream, 7 of 48 corresponding layer-head entries retain their leading norm-associated label, while mean head-norm ANOVA F changes from 24.48 to 18.19. [data] A norm label can reflect unusually low or high category-mean norm.
MLP PR expands at 15 of 16 early and middle layers, then contracts at all 8 late layers
The median Joey/parent MLP participation-ratio ratio is 1.586 in layers 0 to 7 and 1.705 in layers 8 to 15, then 0.287 in layers 16 to 23. [data] PR rises at 7 of 8 early layers and all 8 middle layers, then falls at all 8 late layers. Layer 0 is the early exception, with a Joey/parent PR ratio of 0.682. [data] PR and k90 come from the exported 2,048-row covariance-spectrum subsample.
| Mean-pooled MLP | Parent PR | Joey PR | Parent k90 | Joey k90 | Parent probe | Joey probe |
|---|---|---|---|---|---|---|
| Layer 5 | 123.14 | 209.23 | 764 | 944 | 79.10% | 75.44% |
| Layer 14 | 32.05 | 75.05 | 484 | 816 | 74.59% | 67.33% |
| Layer 19 | 67.47 | 6.47 | 588 | 117 | 78.86% | 69.19% |
| Layer 23 | 13.94 | 1.23 | 96 | 1 | 76.14% | 63.00% |
At layers 5 and 14, higher participation ratio and k90 mean that variance is less concentrated and the 90% tail requires more directions. At layers 19 and 23, both contract: the covariance spectrum is more concentrated and its 90% tail is shorter. [data, inferred] PR measures covariance concentration, not an amount of information lost.
At the final MLP layer, one PC accounts for 90% of the exported subsample's variance and nearest-neighbour label agreement moves from 0.366 to 0.188. [data] This accompanies the stack's largest MLP probe decline, 13.14 percentage points. [data] Joey has 10 rows flagged geometry_dominated at PR below 3, against 2 in the parent; the flag describes concentration, not causal importance.
The length-correlation profile of the leading PCs changes substantially at layer 23; correspondence between the underlying axes has not been established. PC1's correlation with input length moves from 0.854 to 0.279, while PC2 moves from 0.152 to 0.398 and PC3 from 0.083 to 0.333. [data] At layer 14, PC1's length correlation is 0.944 in the parent and 0.931 in Joey. [data] The summaries are not length-residualized, so prompt length remains a competing explanation for some geometry. [inferred]
Health checks report no NaNs or infinities. Both feature tables mark the same 14 layer-0 linattn_qkv coordinates dead, and both sanity tables report those 14 coordinates at each of the two ordinary token positions. [data] These inherited dead coordinates do not account for the selective-count reduction elsewhere in the stack.
18 linear-attention blocks and 6 full-attention blocks are mapped
Both atlases cover 18 linear-attention blocks and 6 full-attention blocks with identical recorded stream shapes and component coverage.
| Block type | Zero-based layers | Exported primary components |
|---|---|---|
| Linear attention | All except the full-attention layers below | gate, up, mlp, linattn_out, linattn_qkv, linattn_z |
| Full attention | 3, 7, 11, 15, 19, 23 | gate, up, mlp, attn, attn_gate, heads, q, k, v |
The 162 component-layer pairs each have last_input and mean_all_real summaries. Six attn_gate streams add 12 sigmoid_last_input and sigmoid_mean_all_real geometry rows. These gate-only sigmoid views are separate from the two ordinary token positions.
last_input is the double newline ending the opened generation prompt. Feature, head, and coactivation tables use the documented primary mean-pooled analysis and omit a position column.
Verdict: decision training accompanies broad backbone reorganization
The result is 5,956 to 1,566 selective MLP entries, 76.04% to 68.89% category-probe accuracy, 0.496 median F-rank correlation, and 11.42% selected-edge retention. [data] Neither tool_use nor data showed the tested increase in single-label selectivity: their selective counts fall to 209 from 869 and 17 from 74, respectively. [data] General category readout declines at every MLP layer, while PR expands at 15 of 16 layers from 0 to 15 and contracts at all 8 layers from 16 to 23. [data] This is consistent with substantial backbone reorganization during full fine-tuning, but the census does not identify the mechanism producing calibrated decisions. [data, inferred]
Limits
- Historical runtime precision, resolved kernels, actual batch partition, and the exact training-parent revision remain unverified. These captures do not isolate a weight-only causal effect. The public capture-code diff changes portal model-size admission, GPU-selection and batching defaults, and Gemma4 MLP-width support; Qwen stream extraction, pooling, tokenization, and runtime-dtype selection are unchanged in that diff. Both captures record the same corpus, template, and token fingerprints, analysis revision, library versions, layer definitions, and stream shapes. [data]
- Joey's configuration names
slot_lmas its readout andblock_bidirectionalas its full-attention override. The inspected capture revision loads genericAutoModelForCausalLM; none of its 44 Python files implements either override or readsmodeling_overrides. The atlas describes the generic causal-LM pass, not Joey's documented OneJev decision-serving configuration. It does not measure decision probabilities, calibration, images, or benchmark performance. [data] - Coordinate associations, probes, neighbourhoods, and coactivation edges are observational evidence, not tested causal circuits. A decision-specific test would use the actual serving mask and readout, decision and abstention prompts, option probabilities, calibration scores, and matched interventions.
Atlases
Qwen3.5-2B · Heretic · Cybersec · Bella Bartender · more under juiceb0xc0de
All six tables are Parquet and in atlas/atlas.sqlite; the queries below use the local SQLite table names.
The selective counts by label; run this query against both pinned atlases for the paired counts:
SELECT top_label, COUNT(*) AS selective_entries
FROM features
WHERE component = 'mlp' AND taxonomy = 'selective'
GROUP BY top_label
ORDER BY top_label;
The geometry and probe curves; run this query against both pinned atlases for the paired values:
SELECT layer, geometry_participation_ratio, geometry_k90,
geometry_twonn, probes_probe_accuracy,
probes_probe_control_accuracy, probes_probe_selectivity,
neighbors_knn_label_agreement
FROM layers
WHERE component = 'mlp' AND position = 'mean_all_real'
ORDER BY layer;
The humor coordinates:
SELECT layer, neuron, top_label, top_auroc, association,
q, neg_log10_q
FROM features
WHERE component = 'mlp'
AND taxonomy = 'selective'
AND top_label = 'humor'
ORDER BY ABS(top_auroc - 0.5) DESC, neg_log10_q DESC,
layer, neuron
LIMIT 20;
The head-label counts; the 7 of 48 retention claim additionally requires joining parent and Joey rows on layer and head:
SELECT norm_top_label, COUNT(*) AS head_entries
FROM per_head
WHERE component = 'heads'
GROUP BY norm_top_label
ORDER BY head_entries DESC, norm_top_label;
Dataset revisions used for every number above:
| Atlas | revision |
|---|---|
| Qwen3.5-2B parent | 6296043f6553ee5c0da5c3d9a686c51f0d5a48c8 |
| Joey-1 | 6d3a26e2fb31f2b39ce5692eb8bcaefffb6e5664 |
- Downloads last month
- 602

