whatisit-nl2sh

Why we included this project

For anyone who lives in a terminal and keeps looking up the exact flags for find, tar, or awk, this tool turns that into a plain-English question. The bundled 941 MB model is a fine-tuned Qwen2.5-Coder-1.5B that runs on CPU through llama.cpp, so answers come back in about a second and nothing leaves the machine, which matters on locked-down boxes where sending shell context to a cloud API is not allowed. The CLI stays minimal: you ask, it prints a command, and nothing runs unless you pass -e and confirm, with destructive operations flagged as DANGER and never auto-executed. If you'd rather use a hosted model, it also works with any OpenAI-compatible endpoint, and the README includes the LoRA training recipe and benchmark numbers against InterCode-ALFA so you can see how it compares to untuned models.