lxask β€” offline Linux command assistant

A 1.5B model fine-tuned to turn plain-English requests into Linux commands. Runs on CPU in about 1.3 GB of memory β€” no GPU, no cloud, no API key β€” and ships with two small LoRA adapters it routes between per request.

Built for lxask, a Rust + llama.cpp CLI that wraps the model with retrieval from the real man pages/tldr corpus, answer checks, and a rule that it never executes a command for you.

Files

file what
lxask-round4-Q4_K_M.gguf the fine-tuned model, Q4_K_M (~986 MB)
restrict-round4.gguf llama.cpp LoRA adapter: ruled-out programs ("without lsof") and proxy contexts
precision-round4.gguf llama.cpp LoRA adapter: near-miss discipline β€” exact filenames, quoting, right flag, leaves keep-files alone
adapters/restrict/ the same adapter in PEFT safetensors format (rank 16)
adapters/precision/ the same adapter in PEFT safetensors format (rank 16)

The adapters are routed, not stacked: restriction/proxy questions get restrict, everything else gets precision. Both GGUF adapters apply to the quantized model directly in llama.cpp.

Measured (held-out sets, commands executed in a sandbox)

  • Strict execution success on generated file/text tasks: 35.0% β†’ 73.8% with the precision adapter (160 tasks)
  • NL2SH-ALFA public benchmark (300 tasks, Docker harness, reward 1.0 only): 62% vs 54% for the untuned Qwen2.5-Coder-1.5B under its strongest prompting (McNemar p = 0.003)
  • Right program on everyday questions: 96–99%; restrictions kept on programs never ruled out in training: 76%

Training

Qwen2.5-Coder-1.5B β†’ SFT on ~39k examples (tldr-pages, NL2Bash, generated requests checked by a validator) β†’ GRPO with commands actually executed in a sandbox as the reward β†’ two rank-16 LoRA adapters trained on contrastive tasks (hard negatives that look right but do the wrong job), each proven to pass reference fixtures and fail no-op/negative variants before training.

Use

Simplest: install lxask β€” it loads these files for you.

curl -fsSLO https://github.com/arxhr007/lxask/releases/latest/download/install.sh
sh install.sh
lxask "count the words in report.txt"   # β†’ wc --words report.txt

Or drive the GGUFs with llama.cpp directly (--lora for an adapter), or load the PEFT adapters onto Qwen/Qwen2.5-Coder-1.5B in transformers.

Limits

A 1.5B model gets details wrong sometimes β€” wrong awk field, wrong uniq sign, occasionally a documentation placeholder left in a path. The companion CLI exists to catch that: it checks every answer against the documentation and marks what it cannot verify. Read the command before you run it.

Downloads last month
135
GGUF
Model size
2B params
Architecture
qwen2
Hardware compatibility
Log In to add your hardware

4-bit

Inference Providers NEW
This model isn't deployed by any Inference Provider. πŸ™‹ Ask for provider support

Model tree for arxhr007/lxask

Adapter
(56)
this model