Identity & provenance
model
SmolLM2-135M-Instruct
type
instruct
maker
Hugging Face (HuggingFaceTB)
released
2024-11
licence
apache-2.0
size
135M
acquire
Hugging Face
the governance surface
Model-documentation dossier
A model card is self-reported, thin, and stale the moment a model is finetuned. The dossier is the read Ardora already does — what this model is, what it tends to do, what its finetune changed, and where it sits among its kin — packaged as a legible, evidenced characterization a governance function can put on file: each claim carrying a replayable witness, each abstain named, each reading stamped with the published ceiling.
Proven this cycle on one family — a documented characterization, a change reading, witnesses, a named marked-empty, and a ceiling stamp. Broader-family documentation is staged, not yet demonstrated. Evidence and documentation material — never certification.
who it's for · what a dossier is
For the function that has to document what it builds on
AI-governance and model-risk teams — the people who owe a legible, evidenced description of what each model is and how it changed, and cannot produce one from a maker's paragraph of intent.
It is the understanding-and-documentation artifact — a legible, third-party, replayable account of what a model is and how it differs from its base. It is not a safety report, an incident record, or an assurance verdict. Ardora answers "what is this?", never "is this safe?"
A dossier is not a new read: it is the one-time read (the disposition characterization, the base↔finetune change reading, the family placement) organised as a documentation-ready field-note that feeds a model card, an internal record, or a transparency disclosure. The wedge is simple — the read Ardora already does leaves behind exactly the legible, evidenced, replayable documentation the transparency obligations reward. Documentation is a byproduct of understanding the model, not a separate project.
what the dossier carries
One model, rendered as a dossier
Every section below is drawn live from this model's field-note in the atlas — the exact shape of the artifact a governance function files.
the model, at a glance
Lineage & kin
base
relationship
instruction-tuned from base (SFT + DPO)
family
size sweep
135M
siblings
5 catalogued kin
Disposition characterization
What the model tends to do — each trait carrying its coarse published confidence band.
instruction-following
markedly more than its base
assistant/chat register
clearly present; the base is a plain text-continuer
Base ↔ finetune change reading
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
witness wit_smollm2-135m_ft · replayable · re-run the finding
abstains — marked-empty, published as loudly as the reading
domain expertise (medical/legal) — insufficient signal to characterize at the published ceiling
reasoning depth — not characterized at the published ceiling for this size
Ceiling stamp — Characterization is descriptive and coverage-bounded; preview engine, single-battery read; see published fidelity ceiling. · engine ardora-core-preview · coverage: single-family, small battery.
where it fits the obligations
Documentation and transparency — the understanding side only
As regulation moves model documentation, provenance, and transparency from good practice toward obligation, a governance function owes a legible, evidenced description of what each model is and how it changed. The dossier fills exactly that gap.
technical documentation
A documented description of what the model is, its characteristics and limitations — the dossier: a witnessed disposition characterization plus published ceilings, organised as documentation material.
provenance of a finetune
Traceability of what a model is and how it changed — the base↔finetune change reading plus lineage placement, each carrying a replayable witness.
legible description
An understandable account of a model’s disposition for downstream users, not just a benchmark score — the legible, calibrated characterization, human-readable field-notes rather than a wall of numbers.
a different bird’s water
The dossier deliberately does not address evaluation-for-risk, adversarial testing, or incident reporting. It documents what a model is and what a finetune changed — the understanding side, not the safety side.
honest limits
What the dossier is, and what it is not
Every characterization is stamped with its published ceiling and its coverage; the dossier states what it could not read as plainly as what it could.
evidence, not certification
The dossier is documentation material a governance function files. It does not certify the model, perform a conformity assessment, or replace counsel or a notified body.
understanding, not safety
It documents what the model is and what a finetune changed. It is not a safety evaluation, an incident record, or a "no hidden behaviour" claim — and no dossier is framed that way.
descriptive & coverage-bounded
Every characterization is descriptive, measured against a published fidelity ceiling, with the confidence stated. The marked-empty list is a first-class section, not a footnote.
staged surface
This cycle demonstrated the exact shape of a dossier on a single family, a small battery — mechanism-proven. Broader-family and larger-model documentation is staged, not yet demonstrated.
No prices or plans on this site — tell the house the models you have to document and we scope the dossier to your case. Every engagement is bespoke. Talk to the house →
See where the documentation fits
The dossier is the understanding-and-documentation artifact. Read how it maps to the EU AI Act's transparency obligations — directionally, and honestly bounded.