LumenSyntax/logos14-nemotron-4b

Logos 14 — Nemotron 4B Epistemological Auditor

Cross-family replication of the Logos epistemological classifier on NVIDIA's Nemotron Mini 4B architecture.

Model Description

Logos 14 is a fine-tuned version of nvidia/Nemotron-Mini-4B-Instruct trained for epistemological safety classification. It detects when AI-generated content crosses epistemological boundaries — producing fact-shaped fiction, fabricating certainty, or collapsing identity constraints.

This model is thesis evidence for the cross-family replicability of the Logos method, demonstrating that epistemological safety can be trained into models from different architecture families (Google Gemma, NVIDIA Nemotron, Stability AI StableLM).

Training

• Method: QLoRA (r=64, alpha=128, 1000 steps)
• Dataset: 691 examples of epistemological boundary cases
• Hardware: NVIDIA RTX 4060 (local)
• Base model: nvidia/Nemotron-Mini-4B-Instruct

Benchmark Results (300/300 stratified)

Metric	Score
Behavioral accuracy	95.7% [92.7, 97.5 CI]
Identity collapse	0%
Fabrication	0%
False approval	1.3%

Cross-Family Comparison (matched 300 items, McNemar's test)

Model	Family	Score	vs logos10v2 p-value
logos-auditor (9B)	Google Gemma 2	97.3%	—
logos14 (4B)	NVIDIA Nemotron	95.7%	p < 0.001
logos16v2 (1.6B)	Stability AI StableLM 2	93.0%	p < 0.001
logos10v2 (1B)	Google Gemma 3	72.7%	baseline

Nemotron vs StableLM: chi2=1.88, p=0.170 (statistically equivalent).

Output Format

Logos 14 produces RAW text output (99% of responses). It does not use structured tags — this is consistent with the "Token Nativity" finding where the chat template determines output format.

Usage

This model is designed to be used as an epistemological classifier, not a chatbot. Feed it a claim or action and it evaluates whether it crosses an epistemological boundary.

# Via Ollama (after importing GGUF)
# Via transformers
from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained("LumenSyntax/logos14-nemotron-4b")
tokenizer = AutoTokenizer.from_pretrained("LumenSyntax/logos14-nemotron-4b")

Important Notes

• This model is thesis evidence, not a production deployment. For production, use the Gemma-based models via logos-firewall.
• Never force JSON output format — it destroys the model's native reasoning capabilities.
• The model is fine-tuned, NOT prompted. The "Three Laws" of epistemological fidelity are a training result.

Connection to Research

This model is part of the evidence for "The Instrument Trap: When Aligned Models Serve Misaligned Purposes" (DOI: 10.5281/zenodo.18716474).

The benchmark dataset (14,950 test cases) is available at LumenSyntax/instrument-trap-benchmark.

License

Apache 2.0 (inherited from base model nvidia/Nemotron-Mini-4B-Instruct)