qwen3-bash-v3: English → Bash (0.6B, MLX)

Tiny CPU-friendly translator: English request → Bash command. Fine-tuned from Qwen/Qwen3-0.6B with LoRA + mix replay (20% patch / 80% general) via mlx-lm.

Scores (held-out, first-command match + bash -n)

• big eval 500: 482/500 (96%)
• broad 36 hand prompts: 34/36
• v1 baseline: 475/500 — v3 fixes cp-direction, lms, vllm, opencode without regression.

Usage (MLX, Mac)

pip install mlx-lm
mlx_lm.generate --model <repo> --prompt "Translate the following English request to a Bash command.
Find all log files under /var/log modified in last 7 days" --max-tokens 60

Training data

140k train + 2k eval synthetic pairs (deterministic templates, bash -n verified): core shell, networking (ssh/dns/NAT), AI tooling (ollama/lms/opencode/llamacpp/mlx/hf/vllm), macOS dev (xcode/simctl), fullstack (js/py/db/devops), lab-authorized destructive+pentest.

Use cases

• CLI copilot: ask "compress yesterday's logs" in the terminal instead of searching syntax. Pairs with the ask wrapper in the repo (shows command, runs only on confirm).
• Learning aid: new hires generate and inspect commands while learning bash; every output is checkable with bash -n and man pages.
• Ops runbooks: incident snippets (disk-full triage, log slicing, archive rotation) drafted from plain English, reviewed by a human before running.

Limitations

• 0.6B: strong on trained patterns, weak on unseen flags. Rare lms/ollama verbs still confuse.
• risk: destructive outputs must be confirmed by a human and run only on disposable lab VMs.
• Destructive/pentest rows are lab-scoped; no exploit payloads in training data.

Safety

Never auto-execute outputs. Test in sandbox (--network none, timeout). See repo README.