Trained a 1.5B to write shell commands so I'd stop googling tar flags. Runs on a laptop CPU

I've been googling "tar extract gz" for about ten years. Finally did something about it. It also works with git, docker, curl, and other everyday tools and tasks.

Its a fine-tuned Qwen2.5-Coder-1.5B on 125k natural-language/command pairs, merged and quantized to Q4_K_M. 941MB, runs through llama.cpp. On my laptop (i5-11320H, 4 threads): 31.9 tok/s, 0.59s median per query, 1.6GB RAM.

Scores 0.620 on InterCode-ALFA. Untuned Qwen2.5-Coder-7B gets 0.613, GPT-4o gets 0.73. Not frontier, but it's roughly a 7B's answer at a quarter the parameters, on a CPU.

There's also a static safety checker with 304 regression cases, because it will absolutely write a command that wipes your root if you ask it to:

Code:

If you want to poke holes in the method or you've got ideas, comment or open a PR. A ⭐ on Github helps and motivates if you find it useful.