Harshell - Natural Language to macOS Shell Commands

Harshell is a fine-tuned Qwen 2.5 1.5B model that converts natural language into macOS shell commands. It returns only the command — no explanations, no markdown, just the shell command you need.

Model Details

Property	Value
Base Model	Qwen 2.5 1.5B
Fine-tuning	LoRA (rank 8, 1000 iterations)
Quantization	Q8_0 GGUF
File Size	~1.5 GB
License	Apache 2.0

Quick Start with Ollama

1. Download the GGUF and Modelfile from this repo
2. Create the model:

ollama create harshell -f Modelfile

3. Run it:

ollama run harshell "list all pdf files in my downloads folder"

Example Usage

Input	Output
list all pdf files in downloads	find ~/Downloads -name "*.pdf"
show disk usage of current folder	du -sh .
kill the process on port 3000	lsof -ti:3000 | xargs kill
compress this folder into a zip	zip -r archive.zip .
show my ip address	ifconfig | grep "inet " | grep -v 127.0.0.1

System Prompt

The model uses this system prompt:

You are a macOS terminal assistant. Convert natural language into safe shell commands. Return only the command, nothing else.

Ollama Modelfile

The included Modelfile configures:

• Temperature: 0.3 (low for deterministic command output)
• Top-p: 0.9
• Max tokens: 128
• Chat template: ChatML format (<|im_start|> / <|im_end|>)

Training Details

• Method: LoRA (Low-Rank Adaptation)
• LoRA Rank: 8
• Training iterations: 1000
• Base model: Qwen/Qwen2.5-1.5B
• Dataset: Curated natural language → macOS shell command pairs
• Quantization: Converted to GGUF Q8_0 using llama.cpp

Files

• harsh-shell-q8_0.gguf — The quantized model (Q8_0, ~1.5GB)
• Modelfile — Ollama configuration file

Limitations

• Optimized for macOS commands; Linux/Windows commands may be less accurate
• Best for single-line commands; complex multi-line scripts may not generate correctly
• Always review generated commands before running them, especially destructive operations (rm, mv, etc.)