CoolFace
Apppublic

AryaYT/nl2shell-demo

sourceHugging Facemitupdated 7mo agoView on Hugging Face
2likes
App README

NL2Shell — Natural Language to Shell Commands

Demo for AryaYT/nl2shell-0.8b, an 800M-parameter model that converts plain English into executable shell commands.

Type a description like:

find all Python files modified in the last 24 hours

and get back:

bash
find . -name '*.py' -mtime -1

No markdown. No explanation. Just the command.

Hardware

This Space runs on the free CPU tier by default (16 GB RAM, 2 vCPU). Generation takes 5–20 seconds on CPU depending on output length. For faster inference:

  • —Duplicate this Space and select a GPU tier (T4 ~$0.40/hr on HF)
  • —Run locally via Ollama (instant, no GPU required on M-series Mac):
bash
  ollama run hf.co/AryaYT/nl2shell-0.8b

Model Details

PropertyValue
Base modelQwen/Qwen3.5-0.8B
Fine-tuning methodQLoRA (4-bit, rank 16, alpha 32)
Training dataNL2Bash + 40 macOS synthetic pairs
Prompt formatChatML
Parameters~800M
GGUF sizesq4km ~400 MB, q8_0 ~650 MB
LicenseMIT

Input Format (ChatML)

The model expects this exact prompt structure at inference time:

<|im_start|>system
You are an expert shell programmer. Given a natural language request, output ONLY the corresponding shell command. No explanations.<|im_end|>
<|im_start|>user
{your natural language request}<|im_end|>
<|im_start|>assistant

The app handles this automatically. If you call the model directly, use this template.

Local Usage

Ollama (recommended)

bash
# Pull and run — downloads q4_k_m GGUF (~400 MB)
ollama run hf.co/AryaYT/nl2shell-0.8b

# One-shot from the command line
ollama run hf.co/AryaYT/nl2shell-0.8b "show disk usage of each subdirectory"

Add to your shell config as a function:

bash
nl() {
    ollama run hf.co/AryaYT/nl2shell-0.8b "$*" 2>/dev/null
}
# Usage: nl find all Python files modified today

Python (transformers)

python
from transformers import AutoModelForCausalLM, AutoTokenizer

MODEL_ID = "AryaYT/nl2shell-0.8b"
SYSTEM = (
    "You are an expert shell programmer. Given a natural language request, "
    "output ONLY the corresponding shell command. No explanations."
)

tokenizer = AutoTokenizer.from_pretrained(MODEL_ID)
model = AutoModelForCausalLM.from_pretrained(MODEL_ID, device_map="auto")

def nl2shell(request: str) -> str:
    prompt = (
        f"<|im_start|>system\n{SYSTEM}<|im_end|>\n"
        f"<|im_start|>user\n{request}<|im_end|>\n"
        f"<|im_start|>assistant\n"
    )
    inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
    outputs = model.generate(
        **inputs,
        max_new_tokens=128,
        temperature=0.1,
        do_sample=True,
        pad_token_id=tokenizer.eos_token_id,
    )
    full = tokenizer.decode(outputs[0], skip_special_tokens=False)
    cmd = full.split("<|im_start|>assistant\n")[-1].split("<|im_end|>")[0].strip()
    return cmd

print(nl2shell("show all running Docker containers"))
# docker ps

Safety

Shell commands can be destructive. Always review generated output before running it, especially commands involving rm, kill, sudo, or network operations. This is a research demo.

Source