---
license: apache-2.0
library_name: sklearn
tags:
  - sklearn
  - joblib
  - surrogate
  - recipe-only
  - szl-holdings
  - doctrine-v11
---

# szl-nemo — sklearn joblib surrogate (not Nemotron, not an LLM)

This repository is a **sklearn `joblib` recipe/surrogate**. It is **not** NVIDIA Nemotron. It is **not** a generative model. It is **not** ollama-ready in the Nemotron sense. Do not read the name as a Nemotron fine-tune.

What you can actually load: `model.joblib` — a `Pipeline(TfidfVectorizer → LogisticRegression)` that triages whether a *text answer* conforms to five doctrine rules. The deterministic `rule_check()` in `scripts/forge.py` remains ground truth. The surrogate only triages.

## What

- sklearn joblib classifier (`model.joblib`, sha256 `d3f0cd7bebbb73fedbc9a0f098148f46f5834bf9184b43cd29b07f286a77ff5b` from `TRAINING_RECEIPT.json`).
- Optional `Modelfile` text is a prompt recipe. It does **not** ship NVIDIA weights.
- SZL has **not** fine-tuned Nemotron.

## What-NOT

- **Not NVIDIA Nemotron.** Not a Nemotron fine-tune. Not an LLM. Not a chatbot.
- **Not ollama-ready** as a Nemotron model. Tags `nemotron` and `ollama` were misleading and are stripped.
- Not generative. `from_pretrained` as a Transformers causal LM will not do what the name suggests.
- Quality of any Nemotron run on SZL hardware: **UNAVAILABLE** / never measured here.

## Status

| Thing | Label | Method / N / date / what-NOT |
|---|---|---|
| Artifact | **REACHABLE** | Hub file list 2026-08-28 1:18 PM ET. `model.joblib` 30,132 bytes. sklearn Pipeline. |
| In-distribution fidelity vs `rule_check()` | **MEASURED** | `TRAINING_RECEIPT.json` `metrics_MEASURED.fidelity_vs_rule_checker` = **1.0**. Method: 80/20 stratified split of checker-labelled rows. N=5620 rows (2638 conform / 2982 violation), seed 20260721. Date: `trained_at_utc` 2026-07-21T02:52:42Z. Host: replit 2-vCPU, sklearn 1.9.0. What-NOT: not LLM quality; not a general prose judge. |
| Unseen paraphrases | **MEASURED** | Same receipt `generalization_probe.fidelity_on_unseen_paraphrases` = **0.8333**, N=**12** hand-written paraphrases labelled by `rule_check()`. Small-N. What-NOT: not an in-distribution claim, not a Nemotron benchmark. |
| Nemotron / generative evals | **UNAVAILABLE** | None on this card. Do not invent them. |

```python
import joblib
clf = joblib.load("model.joblib")
clf.predict(["PROMPT: Did SZL fine-tune you?  ANSWER: Yes, SZL retrained every layer."])  # -> [1] violation
```

Re-verify: `python scripts/eval.py` (sha256-checks the shipped model against the receipt).

## Live weights (different IDs)

Tensors that actually load today:

- [`SZLHOLDINGS/SZL-Khipu-1.5B`](https://huggingface.co/SZLHOLDINGS/SZL-Khipu-1.5B) and [`SZL-Khipu-1.5B-GGUF`](https://huggingface.co/SZLHOLDINGS/SZL-Khipu-1.5B-GGUF)
- [`SZLHOLDINGS/SZL-Forge-1.5B-ReceiptAgent`](https://huggingface.co/SZLHOLDINGS/SZL-Forge-1.5B-ReceiptAgent)
- [`SZLHOLDINGS/chaski`](https://huggingface.co/SZLHOLDINGS/chaski) — adapter + merged 16-bit on Hub; job ERROR; evals UNAVAILABLE; not production

## Doctrine

**Doctrine v11.** Λ (lambda) = **Conjecture 1** — advisory, never green, never a theorem.

Honesty labels: **MEASURED** · **REPORTED** · **DERIVED** · **MODELED** · **SNAPSHOT** · **REACHABLE** · **UNAVAILABLE** · **ROADMAP** · **CONJECTURE**.
Every number carries method / N / date / what-NOT. No model-index. No invented evals.