Offline command inference engine for Linux. Fine-tuned Qwen3.5-0.8B (GGUF Q8_0, 774MB) designed to run on low-resource and edge devices with no GPU, no API, and no internet connection required at runtime.

Benchmark: 99/100 on a structured 100-question Linux command evaluation (Ubuntu 22.04, bash, non-root).

Installation

curl -fsSL https://raw.githubusercontent.com/0-Time/INCEPT.sh/main/install.sh | bash

Supports: Debian/Ubuntu, RHEL/Fedora, CentOS, Arch, openSUSE.

Manual Model Setup

bash

# Download model
huggingface-cli download 0Time/INCEPT-SH \
  incept-sh.gguf --local-dir ./models

# Clone and install
git clone https://github.com/0-Time/INCEPT.sh
cd INCEPT.sh
pip install -e ".[cli]"
incept

Usage

bash

# Interactive CLI
incept

# One-shot
incept -c "list all open ports"

# Minimal output (pipe-friendly)
incept -c "find large files" -m

# With model reasoning
incept --think

CLI Commands

Command	Description
`/think on\|off`	Toggle chain-of-thought reasoning
`/context`	Show detected system context
`/help`	List available commands
`/exit`	Exit

Prompt Format

ChatML with a system context line:

<|im_start|>system
ubuntu 22.04 bash non-root
<|im_end|>
<|im_start|>user
{natural language query}
<|im_end|>
<|im_start|>assistant
<think>
</think>

Inference temperature: 0.0 (greedy decoding).

Training

Parameter	Value
Base model	Qwen/Qwen3.5-0.8B
Training method	Supervised fine-tuning (LoRA, rank 16)
Training examples	79,264 (SFT) + 11,306 (pipe refinement)
Learning rate	5×10⁻⁵
Quantization	Q8_0 (774MB)
Supported distros	Ubuntu, Debian, RHEL, Arch, Fedora, CentOS
Training hardware	Apple M4 Mac mini, 32GB unified RAM

Safety

Prompt injection detection (exact-phrase matching)
Catastrophic pattern blocking (rm -rf /, fork bombs, pipe-to-shell, etc.)
Risk classification: SAFE / CAUTION / DANGEROUS / BLOCKED
Zero outbound traffic at runtime

Requirements

Linux x86_64 / aarch64
Python 3.11+
llama-server on PATH
~1GB RAM at runtime

License

Apache License 2.0

0Time/INCEPT-SH

INCEPT.sh