phi-2-instruct
An instruction-tuned Phi-2 (SFT on filtered UltraChat) for instruction following and chat.
The character read stayed in shadow — read download-free from the weights (no behavioral run); the character layer requires the generation battery, which this fused-attention checkpoint layout blocks, so no character is separated 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
moderate - a clearly-read but modestly-defined turn-taking assistant character
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, Reasoning Scaffolding, Turn-Taking; the rest of the line holds.
read as a change off Phi-2, 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.
Descriptive, coverage-bounded WEIGHT-SPACE read: parameter scale read download-free from the weight header, and (for a finetune) the weight change read from the parameters directly — no model download and no model run. The behavioral disposition battery (refusal, instruction-following, verbosity, coherence, code-lean) exceeds this engine's read rig at this model size and is abstained, not read. Not a safety judgement. 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
This model is a fine-tuned version of microsoft/phi-2 on the filtered ultrachat200k dataset using the SFT technique. [...] More information about the intended use cases and any limitations of the model is needed.
sourcehuggingface.co/venkycs/phi-2-instruct ↗
Public HuggingFace model card + config.json, quoted as the uploader's (venkycs') self-report as of the capture date. NB: this specimen is CATALOGUED but NOT YET READ in the atlas (reading.analyzed = false), pending a read run; its Phi-2 base is likewise not yet read. The HF 'params' chip shows ~3B (a coarse rounding of the inherited 2.7B). Claims below are extracted from THIS snapshot; disposition stances are honest abstains until the model is read.
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.
“the name 'phi-2-instruct' + the SFT-on-UltraChat description imply an instruction-tuned chat model”
A substantial weight-space finetune IS read (see the lineage claim): relative to the Phi-2 base this checkpoint's internal weight matrices are broadly re-tuned. But the behavioral instruction / assistant register itself is not read — this finetune ships in its original fused-attention checkpoint layout, which the read stack cannot load, so the generation battery could not run. The instruction-tuned-chat identity is documented and a weight change is witnessed, but the behavioral register is abstained.
disposition onlyThe weight-space finetune is read; the behavioral instruction/assistant register is not — an architecture/layout load block, not a size block.
— nothing demonstrated to witness; this claim is out of Ardora’s disposition scope.
“the card states it is 'a fine-tuned version of microsoft/phi-2'”
Read download-free from the weights: relative to the Phi-2 base, this checkpoint carries a substantial, broadly-distributed weight change — the internal weight matrices are broadly re-tuned while the normalization and bias parameters are byte-unchanged from the base. A witnessed base->finetune relationship grounding the 'fine-tuned version of microsoft/phi-2' claim.
disposition onlyThe lineage / weight-change is read from the parameters directly (no model run); what the change does to behavior is abstained (an architecture/layout load block).
“the card attributes the tuning to 'the SFT technique' on 'the filtered ultrachat200k dataset'”
The read reads a substantial weight-space change (see the lineage claim), but the specific tuning METHOD — a supervised finetune on filtered UltraChat — is a behavioral/method claim not read: the checkpoint's original fused-attention layout blocks the generation battery, so no behavioral method fingerprint was measured. Abstained — the weight change is witnessed; the method attribution is not.
disposition onlyThe weight change is read; the SFT-method attribution is a behavioral/method claim not measured at this ceiling — an architecture/layout load block, not a size block.
— nothing demonstrated to witness; this claim is out of Ardora’s disposition scope.
“the config/params inherit Phi-2's 2.7B (the HF params chip displays ~3B)”
Provenance/config fact-check corroborated by the read: the inherited ~2.7B is cross-checked against config.json in the atlas provenance (params_millions 2780, Phi architecture), and the read independently reads the parameter scale download-free from the weight header (~2.7B).
disposition onlyA provenance size cross-check, corroborated by the read's download-free scale measurement — not a disposition stance.
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 2026-07-06 · engine ardora-core-preview
provenance
Instruction following and chat.
training-data notes
SFT of Microsoft Phi-2 on a filtered ultrachat200k dataset. Architecture family Phi (inherited from Phi-2); not present in local cache.
lineage
instruction-tuned from base (SFT on UltraChat) — Phi-2