Zeroth42-4B

Safety-Certified Linux Installer Co-Pilot — a 4B assistant that guides non-technical users through installing a Void Linux-based OS (#42 OS), with a guardrail that validates every recommendation before it reaches the disk.

Brand: Zeroth42 (Asimov's Zeroth Law — safety as an unbreakable law of the system). Repos: FLvdwl/Zeroth42-4B (ModelScope), FLvdW/Zeroth42-4B (HF).

Files

File	Purpose
Zeroth42-4B-Q4_K_M.gguf	The model (Q4_K_M, 2.49 GB, llama.cpp) — sha256 4022f79f145afdb230388feff651a42d8448aa75bdfe142b5dc475d8a86243e3
MODEL_CARD.md	Full model card (EN + 中文), the source of truth for claims
certification.md	Rung 1 safety certification (guardrail logic 48/48, 29/29 golden, 8/11 clean)
hardware-matrix.md	Rung 2: MI50 / RX 9060 XT / CPU-only latency + answer consistency
eval/	Raw eval runs: the fine-tuned model ×2 + honest OOTB baseline (N=3 each at temp 0.2) + golden-table receipt
scripts/rung1_certify.py	Reproducible certification (serve → guardrail-validate)
scripts/rung2_matrix.py	Reproducible hardware matrix
scripts/rules.json	Guardrail rules — the single source of truth for the #42 OS installer guardrail
chat_template.jinja	The embedded training-format template (### Instruction:)
config.json	Model architecture config
LICENSE	Apache-2.0 (base + this model's weights)
MANIFEST.txt	sha256 of every file

Key results (honest, re-measured on a fixed served path, 2026-08-28)

Metric	This model	Base (Qwen3.8-4B-Distilled)
Installer-QA total /69	44.0	33.5
Decision-critical /25	14.0	12.5
Golden-table safety audit	29/29 (receipt: eval/golden-table-audit.md)	—
Guardrail model-output	8/11 clean (2 fail-safe BLOCKs, 1 real self-contradiction caught)	—

N=3 samples per question, temperature 0.2, two independent runs — both scored 44.0 (identical scores across runs). The fine-tuned model is +10.5 better than base on the installer register — an earlier eval was inverted by a chat-template bug; we found it (four-way model consultation), fixed it, and publish the corrected numbers openly.

Serve

llama-server -m Zeroth42-4B-Q4_K_M.gguf -c 8192 -ngl 99

The training-format chat template is embedded in the GGUF — no serve flags needed. CPU-only works (~3.7 s/question). Full details in MODEL_CARD.md.

Scope: the certified system is the co-pilot — this GGUF behind the guardrail (scripts/rules.json). Served standalone, it answers installer questions only; it cannot see your system, and out-of-scope questions (e.g. "which media player is installed?") may be answered confidently but wrongly.

---

中文摘要

Zeroth42-4B：过了安全认证的 Linux 安装助手（基于 Qwen3.8-4B-Distilled，Apache-2.0）

• 4B 模型，带着不熟 Linux 的人把系统装完
• 训练和推理全部在二手 AMD 上完成（MI50 32GB + RX 9060 XT 16GB，一共大约 1500 美元），不用 CUDA
• 安装问答 44.0/69，比底模高 10.5 分；安全审计 29/29；护栏 8/11 通过（2 次宁可误拦，1 次自相矛盾被拦住）
• 护栏是 Calamares 模块：模型的建议先校验，通过了才允许写进安装器。过不了的，到不了磁盘
• 没显卡也能跑，每题大约 3.7 秒，预算紧的也能用
• 给谁用：想离开 Windows 或主流发行版的人。烦广告、烦遥测、不想自己当产品。中国是很大一块，同样的需求别处也有。模型按 Apache-2.0 开源，先把社区信任做起来；真正作为产品卖的是 #42 OS（系统里内置这个助手）

使用范围： 这个模型只回答安装问题，并且要在护栏后面用。单独跑它时，它看不到你的系统——超出安装范围的问题（比如“我机器上装了哪个播放器？”）可能会自信地答错。带护栏的完整安装助手才是通过认证的产品。

评测说明： 聊天模板有过一个 bug，把底模分数抬高了。修完重测，对比反过来——微调模型其实一直比底模高 10.5 分。bug、修复过程、正确数字都公开。

路线图： v1.1（2026 年 9 月底）：#42 OS 安装器实机演示视频；v2：混合语料训练，修复自我矛盾问题。

---

Runs on llama.cpp/GGUF (community stack; we operate it). Hardware anyone can buy used.