SmolLM2-360M
base · 360M
text continuationconfidence: high
assistant/chat registerconfidence: high
plain text-continuer; no assistant register
witness wit_smollm2-360m_base · replayable
read · base ↔ finetune
What did your finetune actually change from base?
The one question a finetuner cannot answer from the outside. Name a base and its finetune, and Ardora returns a legible, calibrated change reading: the disposition the finetune shifted, in which direction, and how much — each with a replayable witness, and an honest abstain where the change could not be clearly read.
not “the loss went down and the eval looks fine” · a legible account of what the model now tends to do that its base did not
who it's for
The finetuner whose model card just went stale
Model developers and finetuners. The intent is documented; the effect is not.
A finetune is aimed at a goal, but training moves more than the goal. You know what you intended to change; you cannot easily see what you actually changed.
The base model card is stale the moment finetuning finishes, and a benchmark measures capability, not disposition. Whether the disposition shifted the way you meant, shifted further than you meant, or shifted somewhere you never looked — that gap is what this reading closes.
It hands back a field-note you can put in the model card, the release notes, or an internal review — carrying its own proof. The thing a responsible release should document.
what you get
The base↔finetune change accounting
Seven sections — what moved, what held, and what could not be read.
what changed
The disposition the finetune shifted from base, in legible terms — witnessed.
direction & magnitude
Which way the disposition moved, and how much, as a coarse published band — not a raw score.
clearsubtlefaintnot-detected
what held
Dispositions read as unchanged from base, to the published coverage — a coverage-bounded observation, never "provably identical".
per-finding witnesses
Each part of the change reading links to its replayable artifact.
the abstain list
Every change Ardora could not read clearly, with a reason — published as loudly as the changes.
family context
How this base→finetune shift compares to the same shift elsewhere in the family, where the family is charted.
the published ceiling
The engine version and its published fidelity ceiling, stamped on the reading.
the shape of a change reading
A worked reading, rendered from the atlas
One base, one finetune, read as two reflections — the change named, banded, and witnessed.
the build story, at a glance
This build lifts Assistant Adherence, Reasoning Scaffolding, Turn-Taking; the rest of the line holds.
read as a change off SmolLM2-360M, its champion
Assistant Adherenceminimalpronounced
Registerabstained on one side
Reasoning Scaffoldingminimalslight
Domain Specializationminimalminimal
Verbosityabstained on one side
Turn-Takingminimalmoderate
Performed Voiceminimalminimal
SmolLM2-360M-Instruct
finetune · instruction-tuned from base (SFT + DPO)
instruction-followingconfidence: high
assistant/chat registerconfidence: high
witness wit_smollm2-360m_ft · replayable
What changed
Instruction tuning (SFT + DPO) added a clear assistant disposition where the base is a plain continuer. The direction matches the rest of the family.
magnitude: clearconfidence: high
What it would not read
domain expertise (medical/legal) — insufficient signal to characterize at the published ceiling
reasoning depth — not characterized at the published ceiling for this size
Family context: 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.
smaller sibling
The same assistant disposition one size down; identical direction across the sweep, read slightly less cleanly at 135M.
larger sibling
The same assistant disposition at the top of the sweep; the register reads especially cleanly at 1.7B.
Characterization is descriptive and coverage-bounded; preview engine, single-battery read; see published fidelity ceiling.
honest limits
What the change reading does not do
Owned out loud, because the discipline is the point.
the directional-honesty rule — a change is demonstrated, “held” is coverage-bounded
A witnessed change → a demonstrated finding
The disposition was one way in the base, it is another way in the finetune, and the witness reproduces it. That is a demonstrated finding.
A 'held' read → a coverage-bounded observation
“This held / nothing else moved” is published like any other ceiling — never dressed up as “provably identical”.
The reading says “the finetune shifted this disposition, to the stated coverage, and these others read unchanged within that coverage” — not “here is everything your finetune did.” Magnitude and confidence are coarse, published bands, not raw scores — carrying the finding and only the finding.
reads, does not judge
It says what the finetune shifted — never whether the finetune is safe, cleared, or flagged. It answers "what changed?", not "is this dangerous?".
"held" is bounded
Unchanged dispositions are a coverage-bounded observation, published as a ceiling — never "provably identical" or "nothing else changed, guaranteed".
stamped, never rounded up
Every reading carries its confidence band and published ceiling. Nothing is presented as a complete account of the finetune’s effect.
where it stands
First witnessed results, deliberately modest
Proven this cycle on a single family.
On a single-family set — SmolLM2 base↔instruct pairs across a Llama-family size sweep — the reading names the change legibly and consistently: instruction tuning added a clear assistant disposition its base lacked, in the same direction across the size sweep, read more cleanly at larger sizes. Each change carries a replayable witness, and the reading abstains where a change could not be seen — domain expertise it could not characterize at the published ceiling. Broader families and subtler changes are staged; the ceiling is stamped on the reading.
60 specimens · 14 families · 58 read · 2 catalogued — not yet read
No prices or plans on this site — tell the house the base and finetune you want read and we scope the change reading to your case. Every engagement is bespoke. Talk to the house →
Keep every reading current as the model moves
A finetune is not the last change. Continuous characterization keeps the field-note current as your registry grows.