zero-nl2cmds-v1: An AI Terminal Assistant Model

This is a fine-tuned version of google/codegemma-2b designed to translate natural language instructions into precise Linux/macOS bash commands. It's the core component for an AI-powered command-line assistant.

Author: Sanjayyy06 Version: 1.0

Model Description

This model takes a simple instruction, like "create a single directory named 'api'", and outputs the corresponding bash command, mkdir api. It has been specifically trained and corrected to handle common command-line tasks with a high degree of literal precision.

Intended Use

This model is intended to be the inference engine for a local command-line interface (CLI) tool. A user can type a command in plain English, and the tool will use this model to generate and execute the shell command.

# Example usage with the transformers library
from transformers import AutoTokenizer, AutoModelForCausalLM

model_name = "Sanjayyy06/zero-nl2cmds-v1"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForCausalLM.from_pretrained(model_name)

prompt = "Instruction: list all files in long format\nOutput:"
inputs = tokenizer(prompt, return_tensors="pt")

# Generate the command
outputs = model.generate(**inputs, max_new_tokens=50)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))
# Expected output: Instruction: list all files in long format
#                  Output: ls -l




#Training Process
-The model was fine-tuned using a multi-stage process to ensure high accuracy and control over its behavior.
-Initial Fine-Tuning: The base google/codegemma-2b model was first fine-tuned on a 5,000-example subset of the AnishJoshi/nl2bash-custom dataset.
-Overfitting Correction: Early tests showed the model began to severely overfit on common patterns (e.g., generating entire project structures for a simple mkdir command).
-Surgical Strike: A high-quality, 500-example "correctional dataset" was created to explicitly re-teach the model the literal meaning of simple commands.
-Final Model: The model from step 1 was then fine-tuned for a short duration on the correctional dataset, resulting in this final version which is both knowledgeable and controllable.


#Citation
If you use this model in your work, please cite it as follows:

Code snippet

@misc{sanjayyy06_zero_nl2cmds_v1,
  author = {Sanjayyy06},
  title = {zero-nl2cmds-v1: A Surgically Corrected CodeGemma Model for NL-to-Bash Translation},
  year = {2025},
  publisher = {Hugging Face},
  journal = {Hugging Face repository},
  howpublished = {\url{[https://huggingface.co/Sanjayyy06/zero-nl2cmds-v1](https://huggingface.co/Sanjayyy06/zero-nl2cmds-v1)}},
}