Network transparency for a local LLM: Qwen3-14B v2 NT — a LoRA that announces and reports every network connection, with two disclosed shortfalls / Trasparenza di rete per un LLM locale: Qwen3-14B v2 NT
LoRA adapter for Qwen3-14B v2 that announces before and reports after every network connection (host, country, direction). Recall 99.7%, country/direction 100%, pre-announcement 99.3% (450 dev runs); identity 98% vs 76%; GSM8K 94.9 vs 94.7; ARC-it 90.5 vs 90.6. Limitations disclosed: IFEval 83.5 vs 83.9 (−0.4 pt, p=0.87, not significant) and leak rate 1.11% (5/450) vs a pre-registered ≤1% gate, 95% Wilson CI [0.48%, 2.57%]. Published as a documented exception. Articles in Italian and English. Weights: https://huggingface.co/odelpa/qwen14-scienziato-v2-nt · Code/eval: https://github.com/usergdp136/qwen14-scienziato-v2-nt
Authors
- Giovanni Del Papa
Publication Details
- Journal
- Zenodo (CERN European Organization for Nuclear Research)
- Published
- 2026-11-04
- DOI
- https://doi.org/10.5281/zenodo.23950985
- Primary Topic
- Adversarial Robustness in Machine Learning
- Type
- preprint