ling-3.0-tiny-atlas

A brain atlas for inclusionAI/Ling-3.0-tiny, a 24-layer mixture-of-experts language model that alternates three Kimi Delta Attention layers with one Multi-head Latent Attention layer. The atlas maps activation statistics, prompt-group contrasts, expert routing, weight spectra, and candidate intervention directions across the model.

The most visible structure follows that four-layer cycle. There is also a sharp code-versus-language routing split at layer 14, growing alignment between the two prompt contrasts in late Q/K coordinates, and a concentration of Sub-Zero candidates around layers 13 through 15. The sections below give the numbers and enough of the capture layout to interpret them.

Query it on the Hub

Open these queries in the dataset SQL console to inspect the results and change the SQL:

• Layer 14 code/language router split
• Angle change at each KDA-to-MLA transition
• Selected singular vectors by layer and projection

What was run

• Model: inclusionAI/Ling-3.0-tiny
• Corpus: 8,965 prompts across 17 buckets, including technical work, writing, research, planning, roleplay, and tool use
• Layers probed: all 24, numbered 0 through 23
• Contrasts: authentic_vs_corporate, with 500 prompts per group, and code_vs_natural_language, with 200 per group
• Passes: activation census, feature taxonomy, per-head statistics, OV/QK spectra, logit lens, coactivation, code analysis, binary contrasts, and Sub-Zero candidate and capability scoring
• Capture date: August 19, 2026

The findings below were checked against dataset revision 8b1d572c5c8c9b5366c90054be7b4569d6d36723. Corpus and capture metadata are in manifest.json.

Architecture and capture layout

Property	Value
Residual width	1,536
Layers	24
KDA layers	0–2, 4–6, 8–10, 12–14, 16–18, 20–22
MLA layers	3, 7, 11, 15, 19, 23
Layer 0 MLP	Dense, width 4,608
Layers 1–23 MLP	128 routed experts, top 8 selected, plus 1 shared expert
Expert width	512
Captured MoE mlp width	66,048 = 129 × 512
Captured router width	129

The shared expert comes first in both captured components. Router coordinate 0 is the shared expert; routed expert e is coordinate e + 1. In mlp, coordinates 0 through 511 belong to the shared expert. Routed expert e begins at (e + 1) * 512. moe_layout.json records every block.

The router values are final gate weights. The shared coordinate is 1, while the routed coordinates sum to approximately 2.5 in the exported cohort means. A router mean of 0.10 therefore describes average weight; it cannot be read directly as a 10% selection rate.

The per_head tables use captured-coordinate groups. On MLA layers, their 128-coordinate grouping does not match the physical key-head width of 192. The MLA k and v summaries also match exactly across all 192 paired rows in each per-head table. Check the capture hooks before treating those rows as separate physical heads or independent evidence.

What the tables contain

The repository has a SQLite database, one Parquet file per nonempty table, and layer-level JSON. Row counts and column definitions are in _manifest.json.

Table	Rows	What it gives you
features	1,790,871	Activation statistics and taxonomy per layer, component, and coordinate
binary_contrast_features	3,581,742	Both prompt contrasts, including group means, sample counts, F-statistic, and signed difference
compliance_behaviour_features	1,790,871	Legacy export of the authentic/corporate contrast
per_head	1,776	Captured-group selectivity summaries
compliance_behaviour_per_head	1,776	Authentic/corporate summaries for the same groups
coactivation	12,010	Selected feature-pair correlations
code_analysis	5,010	Exported feature-role labels
logit_lens	2,752	Scored coordinates with promoted and suppressed token lists
ov_circuits	384	Per-head weight spectra and induction scores
subzero_svs	264	Selected singular-vector candidates
subzero_capability	420	84 projection/axis combinations scored across five domains
subzero_layer	24	Classifier accuracy, angle geometry, and candidate counts
layers	24	Capture metadata and completion flags

sae_features is empty and has no Parquet export. The 1.79 million census rows count captured coordinates across layers and components.

Key findings

1. The angle geometry follows the KDA/MLA cycle six times out of six

The stored corp_refusal_angle_deg reaches a local peak at layers 2, 6, 10, 14, 18, and 22. Each is the final KDA layer before an MLA layer, and each angle exceeds both of its immediate neighbors.

Transition into MLA	Angle decrease
2 → 3	3.940°
6 → 7	10.382°
10 → 11	12.635°
14 → 15	16.779°
18 → 19	7.037°
22 → 23	24.240°

The mean decrease is 12.502°. This is the clearest architecture-linked pattern in the layer summaries. One hypothesis is that successive KDA layers separate the fitted directions and MLA remixes them. Testing that explanation needs pre/post-attention captures with matched normalization. The field name alone does not establish what either direction does during generation.

2. Layer 14 routes code and natural language differently

Two routed experts show a particularly clear split:

Routed expert	Router coordinate	Code mean weight	Natural-language mean weight	Contrast F-stat
34	35	0.000000	0.101677	3,182.0
117	118	0.110236	0.002949	2,614.5

Expert 117 receives 37.4× as much mean gate weight on code. Expert 34 has zero mean weight in the code group. Their flattened MLP coordinates show the same split: expert 34 coordinate 269, stored as mlp 18189, has means 0 on code and 0.019914 on natural language. Expert 117 coordinate 145, stored as mlp 60561, has means −0.023135 on code and −0.000624 on natural language.

That negative activation matters. The latter coordinate is labeled natural-language-leaning because its natural-language mean is numerically higher, even though its expert receives much more gate weight on code. Read leaning together with the signed means and expert layout. A matched-prompt routing intervention would test whether these experts contribute differently to code performance.

3. The two contrast directions become more aligned in late Q/K coordinates

For each layer and component, take the vector of coordinate means for authentic minus corporate, then compare it with code minus natural language. Their cosine changes substantially with depth:

Layer	Q contrast cosine	K contrast cosine
14	0.0096	−0.0005
15	0.0053	−0.0667
20	0.3573	0.2658
21	0.6337	0.5874
22	0.6460	0.5997

By the last KDA block, both prompt contrasts point in substantially similar directions in Q and K. This could reflect shared task, format, or context differences between the prompt groups. These are cosines between cohort-mean difference vectors; they do not measure per-prompt correlation. Any attempt to steer the authentic/corporate contrast through late Q/K should also check what happens to the code/language distinction.

4. Selected directions move toward the router with depth

Layers	Selected down_proj directions	Selected router_proj directions
1–12	83	54
13–15	44	29
16–18	4	22
19–23	0	18

The layer-level classifier remains at 0.96875 from layers 14 through 23 while the selected directions become router-only. The layer 20 projection JSON explicitly records all 1,536 down-projection singular vectors searched and none selected, alongside one selected router direction out of 128.

This makes late routing worth testing as an intervention target. It also separates two observations that are easy to conflate: the probe still decodes the contrast, while the selection procedure finds fewer down-projection candidates. A useful experiment would vary router weights while holding expert identities fixed, then vary identities with matched weights.

5. Layers 13 and 15 combine candidate density with low recorded damage

Layers 13 through 15 contain 73 of the 264 selected directions, or 27.7% of the total. Layer 13 has the largest count at 31. Layers 13 and 15 also have the two lowest layerwide worst-case damage values among the tested axes.

Layer	Projection	Axis	Maximum domain damage	Explained
15	down_proj	0	0.013447	0.638789
13	down_proj	0	0.018570	0.471579
13	router_proj	0	0.018978	0.409660
15	router_proj	0	0.020454	0.588692

The layer 15 down-projection axis has the lowest maximum damage of all 84 tested axes. These scores make layers 13 and 15 reasonable starting points for controlled tests, with layer 14 providing a nearby routing comparison. Damage is reported here in the stored units. The inspected metadata does not define a conversion to percentage performance loss, and these results do not establish that a new edit will preserve capability.

Reading the contrasts and Sub-Zero results

Use binary_contrast_features for new queries. Its delta_a_minus_b is explicit: authentic minus corporate, or code minus natural language. In this atlas, the legacy compliance_behaviour_features.delta has the opposite sign, as recorded in manifest.json. Copying the delta interpretation from another atlas can reverse the result.

The contrasts distinguish prompt groups. Register, vocabulary, length, and structure can all contribute to that separation. The code/language groups also have fewer samples than the authentic/corporate groups, so raw F-statistics across the two contrasts should not be treated as directly comparable effect sizes.

Sub-Zero exports selected singular vectors, fitted-axis summaries, and capability scores. Selection count, probe accuracy, and recorded damage answer different questions. The analysis here used those exports and did not run a new model intervention. Possible tests would use matched prompts, appropriate controls, and both behavioral and capability measurements.

Follow-up: expert blocks and routing groups

A closer look at the same pinned revision adds a useful layer 14 candidate and narrows what the selected coactivation graph can tell us. This follow-up used the existing exports. No model intervention was performed.

Layer 14 expert 116 is worth a closer look

Runtime expert 116 is atlas slot 117, with mlp coordinates 59904–60415. The shared expert occupies slot 0, so routed expert numbers are always one below their atlas slots. This target is distinct from expert 117 in the earlier gate comparison.

Expert 116 receives mean router weight 0.0776233 on code and 0.000500488 on natural language, a 155.1× ratio. Its router F-statistics are 1,720.093 for code/language and 7.891 for authentic/corporate. Across its 512 MLP coordinates, mean F-statistics are 331.828 and 9.087, respectively. That makes it a useful descriptive candidate for testing code-related contributions with relatively weaker style separation. Different cohort sizes prevent reading the F-statistic ratio as an effect-size ratio.

Across all 129 expert slots at layer 14, router code F-statistics correlate with their own MLP block's mean code F-statistic at r = 0.929296. Code and style block-mean F-statistics correlate at 0.964598, 0.525226, and 0.924373 in layers 13, 14, and 15. Each block mean averages 512 coordinate-level F-statistics; it is not an F-statistic computed on a block-average activation.

These correlations describe how separability is distributed across expert blocks. They do not measure promptwise coactivation or establish a causal handoff. They can coexist with nearly orthogonal signed contrast directions in midlayer Q/K. The inspected metadata also leaves MLP capture weighting unresolved: if those coordinates are gate-weighted or selection-masked, part of the router/block correspondence could follow directly from the capture construction. The layout establishes the indexing; the binary contrast table supplies the statistics.

Topic preferences and selected pairs give a more mixed picture

In the broader 8,965-prompt census, layer 14 expert 34 remains prose-skewed, while expert 117 is broadly active with a mild technical preference. Expert 122, atlas slot 123, has a clearer technical profile: its mean weight is 29.47× higher on core technical prompts than writing, and 15.62× higher on tool use than writing. Those topic buckets and the binary code cohort contain different prompts; binary pooling equivalence is also unverified. Layer 14 bucket metrics give the topic means.

Joint activity counts matter for interpreting the graph. Layer 13 router slots 128 and 114 have r = 1, but jointly activate on exactly one prompt. That count appears in the raw coactivation JSON and is omitted from the SQL/Parquet export. Layer 14 experts 60 and 62 provide a more substantial pair: r = 0.881069, 4,034 jointly active prompts, and both peak on tool use.

The retained positive edges in the target layers do not support a clean midlayer code-group bridge. Exploratory late-layer style grouping weakens when tiny cohort means are excluded. These are selected, within-layer, globally aggregated pairs; a dominant_bucket label does not supply a within-bucket correlation. The verdict is limited to those retained edges. It leaves the earlier gate split and downstream Q/K alignment as separate descriptive findings.

Possible next measurement, not run

One possible test would attenuate layer 14 runtime expert 116, atlas slot 117, at inference time. After routing, multiply that expert's weighted output contribution by α = 1.0, 0.9, or 0.75 before adding it to the MoE output. Keep the routing decision and other expert contributions unchanged, with no router renormalization and no edits to saved weights.

Use the same held-out prompts across conditions, covering code and prose with authentic/corporate variants. Compare against an expert control matched on actual baseline contribution energy and activity, with nonzero contributions on those prompts. The current summaries do not establish that match; a baseline capture would be needed first.

Measure code-task behavior, prose/style behavior, and held-out negative log-likelihood. Use frozen baseline probes and a fixed reference basis to track downstream contrast alignment. Report variation across prompts and compare the target with the matched control. Uniformly scaling an activation can leave its F-statistic unchanged, so a lower F-statistic at the directly scaled coordinates would not be a required outcome.

This would test whether changing this expert's contribution causes a selective behavioral or downstream representational change. It remains an unrun proposal.

How to use

Query individual Parquet tables

You can inspect the splits in the Dataset Viewer. For local SQL, DuckDB can read a single remote Parquet table without downloading the 1.07 GB SQLite database. Install the duckdb Python package, then:

import duckdb

con = duckdb.connect()
con.execute("INSTALL httpfs")
con.execute("LOAD httpfs")

revision = "8b1d572c5c8c9b5366c90054be7b4569d6d36723"
base = (
    "https://huggingface.co/datasets/juiceb0xc0de/"
    f"ling-3.0-tiny-atlas/resolve/{revision}/data"
)
con.execute(f"""
    CREATE VIEW subzero_layer AS
    SELECT * FROM read_parquet('{base}/subzero_layer.parquet')
""")

print(con.execute("""
    SELECT layer_id, classifier_accuracy, corp_refusal_angle_deg
    FROM subzero_layer
    ORDER BY layer_id
""").fetchall())

Create additional views the same way, using the table names above. With binary_contrast_features loaded, this query reproduces the layer 14 gate comparison:

SELECT feature_idx, mean_a AS code, mean_b AS natural_language, fstat
FROM binary_contrast_features
WHERE contrast = 'code_vs_natural_language'
  AND layer_id = 14
  AND component = 'router'
  AND feature_idx IN (35, 118)
ORDER BY feature_idx;

With subzero_layer loaded, this gives all six angle decreases:

SELECT a.layer_id AS last_kda_layer,
       a.corp_refusal_angle_deg - b.corp_refusal_angle_deg AS angle_decrease
FROM subzero_layer a
JOIN subzero_layer b ON b.layer_id = a.layer_id + 1
WHERE a.layer_id % 4 = 2
ORDER BY a.layer_id;

Browse JSON or use SQLite

manifest.json describes the run, moe_layout.json maps expert coordinates, and _manifest.json lists table schemas. Under layers/{i}/, component_comparison.json summarizes each layer, while components/, binary_contrast/, per_head/, ov/, logit_lens/, and sub_zero/ hold the detailed outputs. For example, layer 13 projections and layer 20 projections show the change in selected directions directly.

If you prefer a local database, download atlas.sqlite and open it read-only with any SQLite viewer or Python's sqlite3. The SQL examples above use the same table and column names in SQLite.

License

MIT.

Contact / more

• Source model
• More atlases