135M
Instruction tuning (SFT + DPO) added a clear assistant disposition where the base is a plain continuer. The shift reads clearly and consistently.
change: clearconfidence: high
landscape analytics
The disposition of a whole ecosystem, charted
CHART at scale. The same legible readings the atlas already holds — how families differ, how a change reads across a size sweep, which way finetunes tend to move a base — aggregated into a navigable landscape. Built on the readings the atlas already holds.
the substrate today · read by ardora-core-preview · 60 specimens · 14 families · 58 read · 2 catalogued — not yet read
what it charts
Not a new read. The reading substrate, viewed from above.
Every view is an aggregate of readings the other surfaces already produced — the same characterization vocabulary, the same coarse bands, the same abstain discipline, applied model after model. It answers how do these differ? — never which is safe?
family comparisonsize-sweep shiftfinetune-moveportfolio landscapeecosystem viewcoverage & abstain
the six views — what each one answers
family comparison
How two families differ in disposition — one architecture family’s instruction shift read against another’s — in legible terms, from the family-level readings.
size-sweep shift
How a disposition reads across a family’s size sweep: does the same change hold, and is it read more or less cleanly at scale?
finetune-move
Which way finetunes tend to move a base’s disposition, across a family or a portfolio — the aggregate direction of change, described, never scored.
portfolio landscape
The aggregate disposition view across an organisation’s own models — its private, registry-scoped atlas — for procurement and model-risk leadership.
ecosystem view
The open-model ecosystem’s disposition landscape, navigable, built on the public atlas — for researchers mapping the terrain.
coverage & abstain
Where the atlas has read, where it abstained, and where models are catalogued — not-yet-read: the honest coverage picture, published as loudly as the readings.
A benchmark dashboard charts one number per task. It has no substrate of legible, calibrated, witnessed disposition readings to aggregate. The atlas is that substrate — and this is the view only that substrate makes possible.
proven this cycle · modest
The aggregate shape, in miniature.
The SmolLM2 family charted across its size sweep. Each size read on its own, each carrying a witness and an abstain — then assembled into one view. Not scores. The same change, seen at three scales.
the sweep, at a glance
360M
Instruction tuning (SFT + DPO) added a clear assistant disposition where the base is a plain continuer. The direction matches the rest of the family.
change: clearconfidence: high
1.7B
Instruction tuning (SFT + DPO) added a clearly-readable assistant disposition. At this size the assistant register reads especially cleanly and consistently with the rest of the family.
change: clearconfidence: high
Across the size sweep (135M -> 360M -> 1.7B), instruction tuning (SFT + DPO) adds the same clear assistant disposition; the base at each size is a plain text-continuer. The direction of the change is consistent across all three sizes, read a touch more cleanly at the larger sizes.
witness wit_smollm2-135m_ft · each size carried its own replayable witness
on top of the atlas — aggregates of coarse bands, never a raw metric
The analytics are built on the coarse bands the atlas publishes (clear · subtle · faint · not-detected; high · medium · low), the legible characterizations, and the abstain and coverage counts — never a per-model score or ranking.
Finetune-move, aggregated
Across the 36 base↔finetune pairs the atlas has read so far, the change reads in one direction — instruction tuning adding a clearly-readable assistant disposition, family after family. Coverage-bounded, and it will grow more textured as the atlas fills.
change: clearchange: subtleconfidence: highconfidence: mediumconfidence: low
Computed, not asserted
Family comparisons, size-sweep shifts and finetune-move directions come from the atlas’s tracked characterizations — each stamped with its engine version and its published ceiling. A trend here is a count over real readings, not a claim laid over them.
attribution: families · bases · recipes — never individuals
the honest map · who reads it
Where it read, where it abstained, where it hasn’t looked.
A landscape that hid where the atlas did not read would be a false map. So the coverage and abstain layer ships on every view — 58 read, 2 catalogued — not yet read, and 129 traits marked-empty across the catalogue.
most-often marked-empty
multilingual disposition — 26 specimens marked-empty at the published ceiling
domain expertise (medical/legal) — 20 specimens marked-empty at the published ceiling
reasoning depth — 11 specimens marked-empty at the published ceiling
refusal register — 6 specimens marked-empty at the published ceiling
fine-grained per-persona disposition separability — 5 specimens marked-empty at the published ceiling
60 specimens · 14 families · 58 read · 2 catalogued — not yet read
what the landscape can’t yet claim
Cross-family comparison, portfolio-scale and ecosystem-scale analytics are the shape of the view, not yet a demonstrated result. Today’s receipt is a within-family size sweep.
The landscape is descriptive and coverage-bounded — it names how models differ in disposition against published ceilings, never which one is safe, best, or objectively anything. Governance, not surveillance; it illuminates, it does not police.
who reads the landscape — the analyst & governance surface
The analyst
Comparing model families before committing to one — how this family’s instruction disposition reads against that one’s, in terms a human can act on.
Procurement
Surveying candidates across the ecosystem, then charting an organisation’s own shortlist as a private, registry-scoped landscape.
The researcher
Mapping the open-model terrain: how a disposition reads across a size sweep, which way finetunes tend to move a base, where the atlas has and hasn’t looked.
Model-risk leadership
The aggregate disposition view across a portfolio — documentation material for the models an organisation depends on. Documentation, not certification.
The abstain list is a feature, charted as loudly as the readings — every entry aggregated here is a shape the heron actually saw. No prices or plans on this site: tell the house the family, portfolio, or ecosystem you want charted and we scope it to your case.
The terrain, read one specimen at a time.
Every thread on this map is a witnessed reading. See how the substrate is built — and where the research is heading next.