Hugging Face's logo

SZLHOLDINGS
/
szl-nemo

surrogate
recipe-conformance
szl-holdings
doctrine-v11
not-a-checkpoint
Model card Files
xet
 Community
6
szl-nemo / README.md
betterwithage's picture
betterwithage
fix: drop the sklearn declaration β€” no loadable artifact exists here
d0cb736 verified
preview code
|
Raw
History Blame
6.59 kB
metadata 
license: apache-2.0
tags:
  - surrogate
  - recipe-conformance
  - szl-holdings
  - doctrine-v11
  - not-a-checkpoint

NOT A LOADABLE MODEL. This repository contains no weight artifact β€” 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.

Operational (MEASURED laptop-Blackwell)

STATUS: tests PASS. get_kernel UNAVAILABLE. Unsloth/LoRA is the wrong tool. Receipted kernels, not silent CUDA.

Thing Label Method / N / date / what-NOT
tests (PYTHONPATH=torch-ext) PASS MEASURED 2026-08-29T15:54:01Z host betterwithage Windows-10-10.0.26200-SP0. torch 2.10.0+cu128. GPU NVIDIA GeForce RTX 5050 Laptop GPU arch Blackwell. pytest 6 passed in 0.02s. Failed nodes: none. What-NOT: not a leaderboard. torch.compile fullgraph failures on Windows Blackwell (cl is not found) are MEASURED, not hidden.
Kernel Hub get_kernel UNAVAILABLE kernels 0.16.1. Default: get_kernel("SZLHOLDINGS/szl-nemo", revision="main", trust_remote_code=True) β†’ False. backend="cpu" β†’ False. trust_remote_code=False β†’ ValueError (SZLHOLDINGS is not a trusted publisher). repo_type=kernel required (kernels 0.16). What-NOT: not a weight load; do not pickle/joblib.load.
formula-tax ADVISORY locked-8 F1 F4 F7 F11 F12 F18 F19 F22. registry_count=21. Ξ› geomean 0.316227766016838. uniqueness Conjecture 1 (never a theorem).
I1–I8 catalog I1 receipt-chain-continuity; I2 ledger-failure-shape; I3 served-run-has-model; I4 signed-columns-atomic; I5 loop-steps-positive; I6 receipt-ed25519-verify; I7 receipt-columns-consistent; I8 flywheel-lineage. Executed by SZLHOLDINGS/szl-invariants. Statuses never coerced. Ξ› untouched.
CUDA speedup / tokens/s / joules UNAVAILABLE Not claimed. Receipted kernels, not silent CUDA.

GitHub source: szl-holdings/szl-nemo @ 6e2f91181788679622613eed6b53871301c536e6. Artifacts: BENCH.laptop-blackwell.json, OPERATIONAL.json.

from kernels import get_kernel
k = get_kernel("SZLHOLDINGS/szl-nemo", revision="main", trust_remote_code=True)

szl-nemo β€” SOFTWARE/SURROGATE (NOT Nemotron, NOT an LLM, NOT a Kernel Hub CUDA kernel)

STATUS: SOFTWARE Β· NOT TRAINED as an LLM Β· no Nemotron/Unsloth weights Β· 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 β†’ 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 β†’ leave β†’ 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 β€” 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. Ξ› = Conjecture 1.