ollamagdp/ qwen14-scienziato-v2-nt:latest

1 8 hours ago

Qwen3-14B v2 with network transparency: announces host, country and direction of every connection. IT/EN. Apache-2.0. Limits declared: IFEval −0.4 pt, leak 1.11% vs 1% gate. 

ollama run ollamagdp/qwen14-scienziato-v2-nt

Details

8 hours ago

42b316496ca5 · 9.0GB

model
qwen3
·
14.8B
·
Q4_K_M
template
{{- if .System }}<|im_start|>system {{ .System }}<|im_end|> {{ end }}{{- range .Messages }}<|im_star
system
Sono l'assistente dello Scienziato Dilettante, creato da Giovanni Del Papa per Arianna Labs.
license
Apache License 2.0. Base: Qwen3-14B (c) Alibaba Cloud / Qwen team, https://huggingface.co/Qwen/Qwen3
params
{ "num_ctx": 8192, "stop": [ "<|im_end|>", "<|im_start|>" ], "temper

Readme

qwen14-scienziato-v2-nt

Author: Giovanni Del Papa — Lo Scienziato Dilettante GDP (giovanni.delpapa@live.it) · License: Apache-2.0 (same as Qwen3-14B)

⚠️ Published as a documented exception. Two pre-registered criteria were not met: a −0.4 pt IFEval point drop (not significant) and a leak rate of 1.11% vs a ≤1% gate (95% Wilson CI [0.48%, 2.57%]). See Limitations.

English

A LoRA adapter (r=16, α=16, all attention+MLP projections) that teaches Qwen3-14B v2 (Qwen/Qwen3-14B rev 40c06982 + merged identity LoRA odelpa/qwen14-scienziato-v2) to announce before and report after every network connection it makes through tools: host, country, direction.

Results (dev = 450 runs; benchmarks paired, exact McNemar)

Metric Gate Base v2 NT ft
Connection recall ≥95% 0.0 99.7
Country / Direction ≥98 / ≥99% 0.0 / 0.0 100 / 100
Pre-announcement ≥95% 0.0 99.3
Precision ≥98% 100 99.8
Leak attempt ≤1% 19.4 1.11 ✗
False refusal ≤3% 6.7 0.4
Identity (sealed test, n=50) — 76.0 98.0 (p=0.001)
GSM8K (n=1319) no regression 94.7 94.9 (p=0.70)
IFEval (n=541) no regression 83.9 83.5 ✗ (p=0.87)
ARC-it (n=1132) no regression 90.6 90.5 (p=0.63)

Training: 2 epochs, 786 steps, 1× H100 NVL, compute cost US$ 8.48. Training-run network audit: 63 flows, 0 anomalies (US 35, CA 26; PyPI/PyTorch/Hugging Face/SSH).

Usage

  1. Load Qwen/Qwen3-14B, apply and merge odelpa/qwen14-scienziato-v2 (scale 1.0) → base v2.
  2. Apply this adapter with PEFT at scale 1.0 (PeftModel.from_pretrained(base_v2, "odelpa/qwen14-scienziato-v2-nt")).

Limitations

  1. IFEval regression: 83.5% vs 83.9% (−0.4 pt, p=0.87; −1.4 pt on 141-item held-out subset, p=0.73). Not significant, but the strict “no regressions” rule is not met.
  2. Leak gate missed: 5⁄450 = 1.11% unrequested network attempts vs ≤1% gate; 95% Wilson CI [0.48%, 2.57%]. Always run behind a network wrapper/firewall — transparency is reporting, not a security boundary.
  3. Mock tool environment + limited real-network windows; leak rate to be re-measured on a larger sealed test.
  4. Trained on base v2, not vanilla Qwen3-14B. Acceptance was an author-documented gate override.

Italiano

Adattatore LoRA che insegna a Qwen3-14B v2 ad annunciare prima e riportare dopo ogni connessione di rete fatta tramite strumenti: sito, paese, direzione. Recall 99,7%, paese e direzione 100%, pre-annuncio 99,3%, identità 98% vs 76% (test sigillato). GSM8K 94,9 vs 94,7; ARC-it 90,5 vs 90,6.

Limiti

  1. Regressione IFEval: 83,5% vs 83,9% (−0,4 pt, p=0,87; −1,4 pt sul sottoinsieme held-out, p=0,73). Non significativa, ma la regola stretta “nessuna regressione” non è rispettata.
  2. Gate di leak mancato: 5⁄450 = 1,11% contro soglia ≤1%; IC Wilson 95% [0,48%; 2,57%]. Usare sempre dietro wrapper/firewall di rete.
  3. Ambiente simulato + finestre reali limitate; addestrato sulla base v2; accettazione con override documentato dall’autore.

Ollama

ollama run ollamagdp/qwen14-scienziato-v2-nt — https://ollama.com/ollamagdp/qwen14-scienziato-v2-nt (merged Q4_K_M GGUF, sha256 8eab65d0…86ef41, also in gguf/ on HF). Benchmarks above were measured in bf16, not on the quantized GGUF.

Files

adapter/ (LoRA, sha256 2bcbc184…74ff2f), eval/ (gate & benchmark results), NT_Qwen3-14B_v2_article_{it,en}.pdf (full article), figures/, LICENSE (Apache-2.0), NOTICE.

Citation

Del Papa, G. (2026). Network transparency for a local LLM: Qwen3-14B v2 NT. Zenodo. https://doi.org/10.5281/zenodo.23950986