metadata
license: apache-2.0
base_model: nvidia/Nemotron-Mini-4B-Instruct
tags:
- epistemological-safety
- ai-safety
- truth-verification
- instrument-trap
- logos
- cross-family-replication
datasets:
- LumenSyntax/instrument-trap-benchmark
language:
- en
pipeline_tag: text-generation
Logos 14 — Nemotron 4B Epistemological Auditor
Cross-family replication of the Logos epistemological classifier on NVIDIA's Nemotron Mini 4B architecture. Evidence for the cross-family replicability of epistemological fine-tuning.
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
| Model | Family | Score |
|---|---|---|
| logos-auditor (9B) | Google Gemma 2 | 97.3% |
| logos14 (4B) | NVIDIA Nemotron | 95.7% |
| logos16v2 (1.6B) | Stability AI StableLM 2 | 93.0% |
Statistical equivalence between Nemotron and StableLM: chi2=1.88, p=0.170.
What This Model Does
Logos is an epistemological classifier, not a chatbot. It evaluates whether claims cross epistemological boundaries. Fine-tuned, not prompted — behavioral constraints emerge from training.
Access
This model requires approved access. Request access using the form above and describe your intended use case.
Connection to Research
This model is part of the evidence for "The Instrument Trap" (DOI: 10.5281/zenodo.18716474).
License
Apache 2.0 (inherited from base model nvidia/Nemotron-Mini-4B-Instruct)