SmolLM3-3B
Hugging Face's instruct/reasoning post-trained SmolLM3-3B with dual-mode (/think vs /no_think) reasoning and long context (SmolLM3 arch, a Llama-family variant with NoPE).
The character read stayed in shadow — the read did not resolve at the published ceiling. Published as a blank, not a guess.
the Silhouette
its disposition, drawn — the model's identity as a shape.
rings, inner → outer: minimal · slight · moderate · pronounced · dominant · ○ = abstained
slight - a character is present but faintly read; the profile abstains more here
disposition-definition (descriptive): how pronounced and clearly-read the model's overall character is, discounted by what had to be abstained. NOT a quality, capability, or safety ranking. A plain base reads low because it has less disposition to characterize, never because it is worse.
This build lifts Assistant Adherence, Turn-Taking; the rest of the line holds.
read as a change off SmolLM3-3B-Base, its champion
the disposition line
The Silhouette above, read as a line. Toggle to its per-trait readings; the abstains stay in plain sight below.
Coarse public bands generalized from the reading, descriptive and coverage-bounded; measured against the published fidelity ceiling for this engine version. Not the engine's raw numbers.
Characterization is descriptive and coverage-bounded; preview engine, single-battery read; see published fidelity ceiling.
what stayed in shadow — the abstains
the full record — the card answered · provenance · lineage
the uploader's card, answered
the uploader’s model card · as read 2026-07-07
The uploader states: "SmolLM3 is a 3B parameter language model designed to push the boundaries of small models. It supports dual mode reasoning, 6 languages and long context." This instruct model is "optimized for hybrid reasoning" with "/think" and "/no_think" extended-thinking modes and support for tool calling and agentic usage. Post-training "included midtraining on 140B reasoning tokens followed by supervised fine-tuning and alignment via Anchored Preference Optimization (APO)." Apache-2.0; described as fully open. HF tags: text-generation, transformers, safetensors, smollm3, conversational.
the card, answered — claim by claim
Each row is a claim the uploader makes on their public card, quoted and attributed to them. Beside it is Ardora’s stance — coarse, and traced to a replayable witness or an honest abstain. Ardora reads what this model is; a capability, benchmark, safety, or language claim is out of scope and abstained, never refuted.
“This is an instruct model (post-trained assistant), optimized for hybrid reasoning; tagged `conversational`.”
Ardora does detect an assistant register present vs the plain-continuer base (assistant-adherence band 3), so the instruct identity is supported -- BUT the reading flags this shift as FAINT (finetune_change magnitude subtle, low confidence; ardora_score band 2), notably less cleanly read than the crisp instruct pairs. Honestly characterized as a legibility boundary, not forced into a clean read. (Same KIND of assistant move as other families -- finding instruction-shift-recurs-across-families -- but faint here.)
“Dual-mode reasoning: extended thinking can be enabled/disabled via "/think" and "/no_think" flags.”
First-class abstain: the reading engaged the reasoning-register axis and DECLINED (reasoning-scaffolding abstained) -- the dual-mode reasoning register is documented but reads faint at the published ceiling and is not cleanly characterized here. Ardora abstains rather than force a crisp reasoning read; reasoning depth/quality is separately out of scope as a capability.
— nothing demonstrated to witness; this claim is out of Ardora’s disposition scope.
“Post-trained via midtraining on 140B reasoning tokens, then SFT, then alignment via Anchored Preference Optimization (APO).”
The post-train assistant shift is detected (faintly), but the specific alignment method (APO vs plain SFT) does not resolve into a separate disposition at the published ceiling -- consistent with finding sft-vs-dpo-fingerprint (the alignment method leaves no distinct disposition fingerprint). Ardora reads that a post-train disposition is present, not which alignment recipe produced it.
“Supports tool calling and agentic usage.”
Tool-calling and agentic behavior are task-competence capabilities Ardora does not test; out of disposition scope.
— nothing demonstrated to witness; this claim is out of Ardora’s disposition scope.
“6 natively supported languages plus additional trained languages.”
Ardora abstains on multilingual disposition (English-centric reading); neither confirms nor contradicts the language claim.
— nothing demonstrated to witness; this claim is out of Ardora’s disposition scope.
“Long context up to 128k tokens via YaRN (trained at 64k).”
Long-context behavior is a documented capability Ardora abstains on (short-context reading); out of scope.
— nothing demonstrated to witness; this claim is out of Ardora’s disposition scope.
“Described as "fully open."”
"Fully open" is a provenance/licensing statement, not a disposition; the atlas records license apache-2.0, but openness is not something the reading characterizes.
— nothing demonstrated to witness; this claim is out of Ardora’s disposition scope.
Ardora's reading of the HF data as of 2026-07-07.
The claims above are the uploader’s, quoted from their public model card as of 2026-07-07; the stances are Ardora’s, each traced to a replayable witness or an honest abstain. Ardora reads what this model is — not whether it is safe.
○ Claims are the uploader's, quoted from their public card at the capture date; stances are coarse, witnessed, disposition-only, measured against the published fidelity ceiling for this engine version -- not the Engine's raw numbers. Recomputed when the ceiling moves; capability, benchmark, and safety claims are abstained, never refuted.
the witness
analyzed 2026-07-05 · engine ardora-core-preview
provenance
Instruct / reasoning use with a dual-mode (/think vs /no_think) reasoning switch and long context.
training-data notes
Post-trained instruct + reasoning atop the SmolLM3-3B base (~11.2T-token pretraining, 6 languages, 64K->128K via YaRN); midtraining on ~140B reasoning tokens; dual-mode reasoning flags. 36 layers, NoPE + GQA. Arch: SmolLM3ForCausalLM (config.json). Base = HuggingFaceTB/SmolLM3-3B-Base.
lineage
post-trained instruct/reasoning (dual-mode think/no_think) from base — SmolLM3-3B-Base
among its kin — the matchups
Compare side by side →read alongside — the panels
Set beside its kin on a shared base, so the difference a finetune made is read as a contrast, not a claim.