TinyLlama-1.1B-NL2SH-Alpaca-v1

Model Card for TinyLlama-1.1B-NL2SH-Alpaca

This model is a fine-tuned version of unsloth/tinyllama-chat.
It has been fine-tuned on the NL2SH-Alpaca dataset for converting natural language instructions into bash commands.

The model outputs one bash command per instruction, even if multiple alternatives exist in the training dataset.

---

Quick start

from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

# Load the model and tokenizer
model_name = "abandonedmonk/TinyLlama-1.1B-NL2SH-Alpaca"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForCausalLM.from_pretrained(model_name).to("cuda")

# Inference helper
def generate_command(model, tokenizer, instruction, inp=""):
    # build prompt in Alpaca-style
    prompt = f"""Instruction: {instruction}
Input: {inp}
Response:
"""
    inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
    outputs = model.generate(
        **inputs,
        max_new_tokens=100,
        do_sample=False,    # greedy decoding
        temperature=0.0,
        num_return_sequences=1
    )
    generated_text = tokenizer.decode(outputs[0], skip_special_tokens=True)
    # Extract the first line (command) after "Response:"
    # If you want to keep all the commands, just simply return 'generated_text' instead of 'response'
    response = generated_text.strip().split("Response:")[-1].strip().split('\n')[0]
    return response

# Example usage
instruction = "Rename all files with .andnav extension to .tile"
bash_cmd = generate_command(model, tokenizer, instruction)
print("Generated bash command:", bash_cmd)

---

Training procedure

This model was fine-tuned using Supervised Fine-Tuning (SFT) on the NL2SH-Alpaca dataset, which contains natural language instructions paired with shell commands.

• Base model: unsloth/tinyllama-chat
• Dataset: abandonedmonk/NL2SH-ALPACA
• Frameworks: PEFT, Transformers, Unsloth
• Number of epochs: 3
• Batch size / seq length: 4

---

Citations

@misc{vonwerra2022trl,
  title        = {{TRL: Transformer Reinforcement Learning}},
  author       = {Leandro von Werra and Younes Belkada and Lewis Tunstall and Edward Beeching and Tristan Thrush and Nathan Lambert and Shengyi Huang and Kashif Rasul and Quentin Gallou{\'e}dec},
  year         = 2020,
  journal      = {GitHub repository},
  publisher    = {GitHub},
  howpublished = {\url{https://github.com/huggingface/trl}}
}

---

License

This model is released under the MIT License

---

Contributors / Maintainers

• Anshuman Jena – fine-tuner, and maintainer of this model 🐸

Notes

• This model is designed for English instructions only.
• Outputs one command per instruction; alternative commands can be manually handled if desired.
• For reproducibility, set the same seed (3407) during fine-tuning.