SZLHOLDINGS/szl-nemo

NOT A LOADABLE MODEL. This repository contains no weight artifact — no joblib, no npz, no safetensors. It previously declared library_name: sklearn with sklearn and joblib tags, which told the Hub to present it as a loadable scikit-learn model; nothing here can be loaded that way. get_kernel is UNAVAILABLE (returns False), so the kernel path does not resolve either. Both the sklearn declaration and the kernel path have been removed from the metadata rather than left to imply a capability that is absent. The surrogate rules and tests in this repo are real; the checkpoint is not.

S Z L  N E M O

Honesty over SEO: scripts and a receipt, no checkpoint. Do not invent the blob.

STATUS: SOFTWARE · NOT TRAINED as an LLM · no Nemotron/Unsloth weights · joblib UNAVAILABLE.

Canonical GitHub source: szl-holdings/szl-nemo. Hub ID SZLHOLDINGS/szl-nemo is a sklearn recipe-conformance surrogate card. It is not NVIDIA Nemotron. It is not a generative model. It is not a Triton/CUDA kernel. Tags nemotron and ollama were misleading and are stripped. Do not from_pretrained this as an LLM. SZL has not fine-tuned Nemotron and does not republish NVIDIA weights.

Approved GitHub path: szl_nemo.rule_check (stdlib, R1–R5). model.joblib is quarantined.

What it is / is NOT

• IS: scripts/forge.py + scripts/eval.py + TRAINING_RECEIPT.json describing a Pipeline(TfidfVectorizer → LogisticRegression) that triages whether a text answer conforms to five doctrine rules (R1–R5). Deterministic rule_check() in scripts/forge.py remains ground truth. Optional Modelfile is prompt text only.

The cut

We took the idea of recipe-conformance from NVIDIA NeMo and built a tiny sklearn surrogate that triages answers against five doctrine rules. Then we stripped the misleading nemotron tags. Honesty over SEO.

A 10-millisecond 'does this answer violate doctrine?' that CI can run on every card.

Silhouette → leave → SZL

Leader	Take, then tweak
Anthropic	Constitutional classifier, tiny.
NVIDIA	Silhouette of NeMo recipe-conformance. Cut: sklearn, disclosed, not a Nemotron.
Unsloth	No.

Nobody else ships this combination. That is the point of a one-of-one.

Intended use

CI doctrine triage. Retrain from forge.py.

Limitations

• model.joblib not on Hub at snapshot.
• Not Nemotron. Not generative.

Canonical GitHub: szl-holdings/szl-nemo

• NOT: NVIDIA Nemotron 3 Nano 4B. Not ollama-ready Nemotron weights. Not a chatbot. Not a fine-tune. BASE_MODEL_MANIFEST.json is an observation of an upstream Ollama tag (mutable); it is not weights in this repo.

Status

Thing	Label	Method / N / date / what-NOT
model.joblib on Hub	UNAVAILABLE	Hub file list 2026-08-28 ~6:56pm ET. Files on main: .gitattributes, BASE_MODEL_MANIFEST.json, LICENSE, Modelfile, README.md, SZL_ESTATE_MANAGED.json, TRAINING_RECEIPT.json, scripts/eval.py, scripts/forge.py. No model.joblib. Receipt names file model.joblib sha256 d3f0cd7bebbb73fedbc9a0f098148f46f5834bf9184b43cd29b07f286a77ff5b — that blob is not published here. Do not invent it.
Receipt-bound scorer metrics	REPORTED in TRAINING_RECEIPT.json	trained_at_utc 2026-07-21T02:52:42Z, host replit 2-vCPU, sklearn 1.9.0, seed 20260721. N=5620 checker-labelled rows (2638 conform / 2982 violation), 80/20 stratified. fidelity_vs_rule_checker 1.0; unseen paraphrases 0.8333 (N=12). What-NOT: not LLM quality; not a Nemotron benchmark; cannot be replayed from Hub bytes until model.joblib is present.
Nemotron / generative evals	UNAVAILABLE	None on this card. Quality of any Nemotron run on SZL hardware: UNAVAILABLE.
NVIDIA weights	NOT REPUBLISHED	Never copy upstream tensors into this ID.

When model.joblib is actually committed, load with joblib.load("model.joblib") and sha256-check against the receipt. Until then, this ID is scripts + a receipt, not a loadable sklearn artifact.

Apache-2.0 for SZL files here. Upstream Nemotron, if you fetch it yourself, stays under NVIDIA's license. Λ = Conjecture 1.

---

Hub: SZLHOLDINGS/szl-nemo