Generating malicious scripts or exploits.
Executing commands automatically without human verification (Always review the output before running it in a terminal).
Complex scripting in languages other than Shell (unless specified).

Quick Start

You can use the code below to run inference with the model:

python

from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "andreagalasso99/Qwen3.5-4B-NL2SH"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(model_id, device_map="auto")

SYSTEM_PROMPT = """You are a Linux System Automator. \
Your goal is to convert natural language instructions into precise, \
safe Bash commands. Given the following instruction, \
output a JSON object with two keys: 'command' (the string to execute) \
and 'safety_warning' (a string describing eventual risks, or 'none')."""

def generate_command(user_instruction):
    messages = [
        {"role": "system", "content": SYSTEM_PROMPT},
        {"role": "user", "content": user_instruction},
    ]
    
    # Apply ChatML template
    prompt = tokenizer.apply_chat_template(
        messages, 
        tokenize=False, 
        add_generation_prompt=True
    )

    inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
    
    outputs = model.generate(
        **inputs,
        max_new_tokens=128,
        temperature=0.7,
        do_sample=True,
        pad_token_id=tokenizer.pad_token_id,
        eos_token_id=tokenizer.eos_token_id
    )
    
    # Decode and extract only the new tokens
    decoded_output = tokenizer.decode(outputs[0][inputs['input_ids'].shape[1]:], skip_special_tokens=True)
    return decoded_output.strip()

# Example Usage
instruction = "Find all files in /home/user/ that end with .txt and count them."
response = generate_command(instruction)
print(response)

Training Data

The model was trained on the NL2SH-ALPACA dataset ((https://huggingface.co/datasets/abandonedmonk/NL2SH-ALPACA). This dataset follows the Alpaca instruction format:

Instruction: The natural language task.
Input: (Optional) Context for the command.
Output: The corresponding Shell command.

Training Details

Hyperparameters

Epochs: 2
Batch Size: 64
Learning Rate: e.g., 2e-4
Optimizer: AdamW
Weight Quantization: 4-bit (NF4)
Precision: Float16

Tools & Frameworks

PEFT/LoRA
Unsloth
Hugging Face Transformers
TRL

Hardware

Nvidia A100