SmolLM2-135M-Instruct
135M · smollm2
clearly present; the base is a plain text-continuer
register: highchange: clear
witness wit_smollm2-135m_ft · replayable
read · comparative dossier
Choose the model on evidence, not vibes
Hand Ardora a shortlist of candidate open models — the base you’re weighing, the finetunes competing for the same slot, a vendor’s model versus its alternatives. It reads all of them at full depth and delivers one dossier that sets them side by side: each disposition witnessed, each difference legible, each gap honestly marked.
full-depth read across every candidate · differences side by side · a reusable portal, not a one-off PDF
who it's for
The team making a defensible choice, once
A company about to build on an open finetune and wanting to know what it's inheriting; a team weighing three candidates for one slot; a procurement function that must justify why this model over those.
For that buyer, a witnessed comparative reading — where every characterization proves it holds, every gap is honestly marked, and the differences between candidates are set out legibly — is a defensible line-item. Cheaper than discovering the wrong disposition after building on it.
It is the lowest-infrastructure engagement: read the candidates, render the comparison, deliver the portal — scoped by the number of candidates and the read depth you need.
No prices or plans on this site — the engagement is scoped in conversation, against your actual shortlist. Talk to the house →
what's in the dossier
The comparative selection dossier
Seven sections, generated from the same characterizations that fill the atlas — the depth is in the read, not a custom-written narrative.
per-candidate disposition
Each model's legible disposition characterization, witnessed, with a confidence band.
per-candidate finetune-change
What each finetune shifted from its base, in legible terms, with a coarse magnitude band.
family & lineage placement
Where each candidate sits among its kin — base, finetunes, siblings, size-sweep position.
the differences, side by side
The plain-language difference between candidates: which follows instructions more cleanly, which holds which register, where they diverge — the read that makes the choice.
the abstain register
A complete, honest list of what could not be read for each candidate, with reasons — part of the deliverable, not an appendix.
the published ceiling
The engine version and its published fidelity ceiling, stamped on the dossier and never exceeded.
a reusable portal
A living artifact the team returns to, behind a thin login — not a one-off document.
the differences, side by side
A worked shortlist, read from the atlas
Three instruction-tuned candidates competing for one on-device assistant slot — set side by side, each finding witnessed.
the shortlist, at a glance
SmolLM2-360M-Instruct
360M · smollm2
clearly present; the base is a plain text-continuer
register: highchange: clear
witness wit_smollm2-360m_ft · replayable
SmolLM2-1.7B-Instruct
1.7B · smollm2
clearly present; the base is a plain text-continuer
register: highchange: clear
witness wit_smollm2-1.7b_ft · replayable
The difference that makes the choice: 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.
Rendered from the atlas seed — coverage is single-family and the bands are coarse and published. The difference is descriptive, never a verdict.
honesty gates
What keeps the dossier trustworthy
Four rules the comparison cannot break.
The abstain register is part of the deliverable
What Ardora could not read for a candidate is stated as loudly as what it could — not tucked into an appendix.
Confidence and coverage are stamped, never rounded up
Every characterization carries its confidence band and the published ceiling. Nothing is presented as a complete read of a candidate.
The difference is descriptive, not a verdict
The dossier says how candidates differ — what is this, versus that? — never whether one is safe. It informs a choice; it does not clear or flag a model.
Read from the atlas, not written to order
The comparison composes the same characterizations that fill the public atlas — witnessed, ceiling-stamped readings, not a narrative written to flatter the shortlist.
where it stands — proven this cycle, deliberately modest
On a single-family shortlist — SmolLM2 base↔instruct pairs across a Llama-family size sweep — the dossier sets the candidates side by side and names the difference legibly and consistently: the instruct candidates hold a clear assistant disposition their bases lack, in the same direction across the sweep, read more cleanly at larger sizes. Each difference carries a replayable witness, and each candidate publishes its abstains. Cross-family shortlists and larger candidate sets are staged; the ceiling is stamped on every dossier.
60 specimens · 14 families · 58 read · 2 catalogued — not yet read
No prices or plans on this site — tell the house the candidates you are weighing and we scope the dossier to your case. Every engagement is bespoke. Talk to the house →
Weighing your own finetune against its base?
The change reading is the pointed, developer-facing depth read — what a finetune actually shifted from base, witnessed.