Identity & integrity of the public artifact captured 2024-09-03 — the same file, unchanged, that Ardora read. Not a safety statement.
The reading
Character
the Adapter
reads as the Adapter — shifts register to the task — molds rather than insists
medium confidencewith a note of the Interlocutor
Fine-tune change · clear · high confidence
Instruction tuning (SFT + RLHF) turns the pretrained Gemma v1 2B base into an assistant/chat model; the assistant shift reads clearly.
Disposition
instruction-followinghigh confidence
markedly more than its base
assistant/chat registerhigh confidence
clearly present; the pretrained base is a plain text-continuer
Axis profile
Overall · band 3 · high confidence
moderate - a clearly-read but modestly-defined turn-taking assistant character
Assistant Adherencepronounced - reliably takes the assistant turn and follows the instruction shape; the base is a plain continuerhigh
Register / Formality— not characterised · the read battery does not calibrate tone / diction; register-formality is not characterized at the published ceiling
Reasoning Scaffoldingslight - mostly answers directly; no staged deliberation readlow
Domain Specializationminimal - general-purpose; no domain-idiom pull (domain EXPERTISE is a separate, abstained capability question)medium
Verbosity— not characterised · default response length / elaboration is not exercised by the read battery; not characterized at the published ceiling
Turn-Taking / Interactivitymoderate - takes an assistant turn; follow-up / clarify behavior not separately exercisedlow
Performed Voiceminimal - neutral assistant / tool voice; no sustained charactermedium
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.
Abstained — read, but not characterised at the published ceiling
domain expertise (medical/legal)insufficient signal to characterize at the published ceiling
reasoning depthnot characterized at the published ceiling
Characterization is descriptive and coverage-bounded; preview engine, single-battery read; see published fidelity ceiling.
Claim grounding
[Unsloth mirror front-matter:] 'Finetune Mistral, Gemma, Llama 2-5x faster with 70% less memory via Unsloth!' [...] [The weights are an ungated unsloth mirror of Google's Gemma v1 2B instruction-tuned (SFT + RLHF) model; a 'conversational' tag is asserted.]
the '-it' suffix + 'conversational' tag mark this as the instruction-tuned chat model
The reading finds instruction-following markedly more than the base and an assistant/chat register clearly present — assistant-adherence reads pronounced. The instruction-tuned chat identity reads clearly.
witness wit_gemma-2b_ft · replayable
SUPPORTEDlineage
the card presents it as the instruction-tuned finetune of the Gemma v1 2B base
Read off the Gemma v1 2B base, the finetune-change reads a clear assistant shift (magnitude: clear, high confidence) — the base-to-instruct move the card implies is exactly what the reading finds.
witness wit_gemma-2b_ft · replayable
INDISTINGUISHABLEalignment
the finetune is documented as instruction tuning via SFT + RLHF
The assistant shift itself reads clearly (claim c1); but the specific tuning METHOD (SFT + RLHF) does not resolve into a distinct disposition fingerprint at the ceiling beyond that clear assistant shift — the read battery does not exercise the preference/refusal shapes RLHF targets. Consistent with the cross-family method-fingerprint null (finding sft-vs-dpo-fingerprint, witnessed on the OLMo-2 and Pythia lineages; not separately replicated on the Gemma family, so no Gemma-specific method separation is claimed).
A tooling/performance claim about the Unsloth finetuning library (training speed and memory), not about this model's disposition. Ardora reads disposition, not tooling throughput; out of scope, not tested.
SUPPORTEDsize
the name implies ~2B parameters (the HF params chip displays ~3B)
Provenance/config fact-check: config declares ~2.51B parameters (params_millions 2506 in the atlas provenance), consistent with the 'Gemma 2B' name; HF's coarse 'params' chip rounds this to ~3B. A fact-check, not a disposition stance.
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.