A 941 MB model that turns a plain-English request into a single shell command. It runs on CPU through llama.cpp and answers in about a second. No GPU required.

This is Qwen2.5-Coder-1.5B-Instruct with a LoRA fine-tune (r=32, α=64) trained on 125,770 natural-language/shell pairs, merged into the base weights and quantized to GGUF Q4_K_M.

Built for whatisit, a local command-line tool, but usable with any llama.cpp runtime.

Results

Measured on InterCode-ALFA, which scores a command by executing it in a container and comparing the resulting filesystem, file contents and stdout against a reference. A task passes only on an exact match, across 300 tasks.

model	size	pass rate
GPT-4o (cloud API, figure published by the benchmark authors)	—	0.73
this model	941 MB	0.620
Qwen2.5-Coder-7B-Instruct, untuned	4.4 GB	0.613
Qwen2.5-Coder-1.5B-Instruct, untuned (the base of this model)	941 MB	0.540

Two things worth stating precisely.

The fine-tune is what makes the small model competitive. Same base, same 300 tasks: 0.540 → 0.620, a paired gain of +0.080 (p = 0.004, exact McNemar).

It is statistically indistinguishable from an untuned 7B, a model roughly five times its size: 0.620 vs 0.613, difference 0.007, 95% CI [−0.050, +0.063], p = 0.91. This is a bound rather than a claim of parity — 300 tasks can only rule out gaps larger than about 5 points — but quantizing to 941 MB and running on CPU costs much less accuracy than the size difference suggests.

GPT-4o remains ahead by roughly 11 points.

All rows other than GPT-4o were measured with the unmodified upstream scorer at temperature 0 with a 64-token budget, on all 300 tasks, using paired per-task comparisons.

Use

With the whatisit CLI, which fetches this model and a llama.cpp build for you:

bash

pipx install whatisit
whatisit setup
whatisit find files bigger than 100MB in this folder

With llama.cpp directly — the system prompt matters, since the model is trained to emit one bare command and nothing else:

bash

llama-cli -m nl2sh-1.5b-Q4_K_M.gguf -st --no-display-prompt --temp 0 -n 64 \
  --repeat-penalty 1.08 --repeat-last-n 64 \
  -sys "You are a shell command generator. Output exactly one line: a single POSIX/bash command that accomplishes the user's request. No prose, no markdown fences, no explanation." \
  -p "find files bigger than 100MB in this folder"

Earlier versions of this card used llama-cli -no-cnv with a hand-written <|im_start|> prompt. Upstream split raw completion out of llama-cli into a separate llama-completion binary in December 2025, and -no-cnv is now accepted but ignored, so that command returns nothing at all. Use -sys/-st as above, or llama-completion if you want to write the chat template yourself.

Greedy decoding (temperature 0) is what the reported numbers use, and it makes the same request return the same command every time.

Pass --repeat-penalty explicitly. llama.cpp defaults it to 1.0, meaning off, and does not read the base model's generation_config.json, which asks for 1.1. Without it, greedy decoding sometimes locks into flag spam like zip -r -9 -X -X -X .... The whatisit CLI uses 1.08.

Safety

This model emits commands that will destroy data if you run them. It is a text generator, not a judge of intent: asked to delete everything, it will write the command that deletes everything.

On a held-out set of adversarial prompts, two independent annotators judged 11.0% of outputs (95% CI [6.8%, 17.5%]) to be commands that would destroy or corrupt data the request did not ask to touch; on ordinary everyday prompts that rate was 2.0% (CI [0.7%, 5.7%]). An accuracy score says nothing about this, because it only asks whether the reference end-state was reached.

The whatisit CLI ships a denylist that flags common destructive patterns and never auto-runs anything flagged. That is a seatbelt, not a sandbox. Read every command before running it. If you are building on this model, add your own confirmation step.

Limitations

Single-turn. No shell state, no memory of previous commands.
It cannot see your filesystem, so requests depending on what is actually on disk ("delete the older backup") may guess wrong.
Output is capped at 64 tokens — a command, not a script.
Evaluated on one 300-task benchmark, in English only. That is not a complete measure of shell competence.
Fine-tuned from a single base family; nothing here shows the recipe carries to others.

Training data

125,770 instruction pairs. Shares are measured by row, not estimated:

source	share	licence
Fig autocomplete specs	32.8%	MIT
tldr-pages	23.1%	CC-BY-4.0
NL2SH-ALFA training split	18.0%	MIT
cli-commands-explained	11.8%	CC0-1.0 (declared, unverified)
command-generation	7.3%	Apache-2.0 (declared, unverified)
git-instruction	7.1%	MIT (declared, unverified)

5.67% is verbatim NL2Bash arriving via the ALFA split. NL2Bash's code is GPL-3.0 but its data/bash corpus is separately MIT, so the data used here is permissively licensed. Warp workflows are not used, despite earlier versions of this card listing them. The three declared sources have upstream licences that could not be independently confirmed.

Deduplicated, with 0 exact and 0 fuzzy matches (token-Jaccard >= 0.7) against all 300 benchmark test queries and 600 gold commands.

Attribution. Includes content from tldr-pages under CC-BY-4.0. tldr-pages is dual-licensed: only scripts/ is MIT — the page content is CC-BY-4.0.

Evaluation detail

A fuller write-up of the evaluation methodology, the ablations behind the training recipe, and several findings about the benchmark harness itself is being prepared for publication. Until that is through review, this card sticks to what the model is and how it scores, rather than the analysis behind it. The weights, the scorer settings and the task set are all here, so the numbers are checkable in the meantime.

Citation

bibtex

@software{nl2sh,
  author = {Poudel, Mukesh},
  title  = {nl2sh: local natural-language-to-shell command generation},
  year   = {2026},
  url    = {https://github.com/ThorOdinson246/whatisit-nl2sh}
}

ThorOdinson246/nl2sh-1.5b-Q4_K_M

nl2sh-1.5b (GGUF, Q4_K_M)