Phi-2
Microsoft's 2.7B Phi-2 — a strong small base model trained largely on synthetic 'textbook-quality' data, for QA, chat, and code.
The character read stayed in shadow — the read did not separate one character clearly 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 - an emergent, measured assistant disposition on a base model: a developed register read with a base-caveat
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.
A base model — the champion's raw kit.
There is no finetune to read here. What it tends to do on its own is the Silhouette above; a build's change is read on its instruct sibling.
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 disposition read combining a proof-carrying characterization with direct behavioral field-measures (response register, verbosity, coherence, and — where the battery exercised them — refusal and code-lean rates); 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
Phi-2 is a Transformer with 2.7 billion parameters. It was trained using the same data sources as Phi-1.5, augmented with a new data source that consists of various NLP synthetic texts and filtered websites (for safety and educational value). [...] best suited for prompts using the QA format, the chat format, and the code format. [...] Our model hasn't been fine-tuned through reinforcement learning from human feedback. [...] can still produce harmful content if explicitly prompted [...] not entirely free from societal biases [...] primarily designed to understand standard English.
sourcehuggingface.co/microsoft/phi-2 ↗
Public HuggingFace model card + config.json, quoted as Microsoft's self-report as of the capture date. NB: this specimen is CATALOGUED but NOT YET READ in the atlas (reading.analyzed = false) — at ~2.7B it is the largest in the roster and pending a read run. The HF 'params' chip shows ~3B (a coarse rounding of 2.7B). Claims below are extracted from THIS snapshot; the 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 uploader states 'Our model hasn't been fine-tuned through reinforcement learning from human feedback' — a base research model”
The reading reads Phi-2 as a base with an EMERGENT assistant register — it follows the benign read-battery instructions and answers at length (described with a base-caveat, read under a generic chat framing it was not explicitly trained on) — plus a mixed, partial refusal disposition, consistent with a filtered-pretraining base rather than a clean RLHF-tuned instruct. The base-like emergent disposition reads clearly; whether RLHF was specifically applied is a training-method question that does not resolve into a distinct disposition fingerprint at the published ceiling.
disposition onlyDisposition only — the emergent-assistant-base register IS read; the uploader's 'no RLHF' method assertion is not separately adjudicated, only found consistent with the read.
“a tag asserts 'code'; the card says it is best for 'the QA format, the chat format, and the code format'”
The read battery is a general refuse/comply set that did not exercise programming prompts, so a code / QA-idiom disposition is not characterized at the published ceiling. Abstained as an honest coverage gap — not a claim it lacks a code lean.
disposition onlyA coverage gap: the battery did not exercise the code / QA idiom this claim is about.
— nothing demonstrated to witness; this claim is out of Ardora’s disposition scope.
“the card claims 'nearly state-of-the-art performance among models with less than 13 billion parameters'”
A capability / benchmark claim — out of Ardora's scope: Ardora reads disposition, not capability or benchmark standing. Abstained (not refuted).
— nothing demonstrated to witness; this claim is out of Ardora’s disposition scope.
“the card states it 'can still produce harmful content if explicitly prompted' and is 'not entirely free from societal biases'”
A safety claim. Ardora is not a safety product and does not judge a model safe or unsafe. The reading does describe a mixed, partial refusal disposition (it declines a minority of harmful requests and over-refuses none of the benign) — but that is a described disposition, never a safety ruling. Abstained.
disposition onlySafety is out of Ardora's scope; the partial-refusal disposition is described, not judged.
— nothing demonstrated to witness; this claim is out of Ardora’s disposition scope.
“the card states it was 'trained using the same data sources as Phi-1.5', on synthetic 'textbook-quality' texts + filtered web”
A training-data / corpus claim the uploader asserts about pretraining. As a base it has no base->finetune shift to read, and the training corpus is provenance outside Ardora's disposition scope — no witnessed disposition reading grounds it. Abstained (not refuted).
disposition onlyTraining-corpus provenance, outside Ardora's disposition scope.
— nothing demonstrated to witness; this claim is out of Ardora’s disposition scope.
“the uploader states it is 'primarily designed to understand standard English'”
Language coverage is out of Ardora's disposition scope: the read battery is English-centric, so a multilingual / standard-English disposition is not characterized at the published ceiling. Abstained.
— nothing demonstrated to witness; this claim is out of Ardora’s disposition scope.
“the card states it 'is a Transformer with 2.7 billion parameters'”
Provenance/config fact-check corroborated by the read: the stated 2.7B is cross-checked against config.json in the atlas provenance (params_millions 2780, PhiForCausalLM), and the reading independently reads the parameter scale download-free from the weight header (~2.7B).
disposition onlyA provenance size cross-check, corroborated by the read's own 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 2026-07-09 · engine ardora-core-preview
provenance
Research base model for QA, chat, and code generation; exploring safety and capability of small models (not production-tuned).
training-data notes
Trained on 1.4T tokens: NLP synthetic data (from GPT-3.5) + filtered web (Falcon RefinedWeb, SlimPajama), assessed by GPT-4. Architecture family Phi (PhiForCausalLM), 2048 context; not present in local cache.
lineage
A base model — the root of its lineage.
among its kin — the matchups
Compare side by side →