starcoder-cli 🌟

This is an example of a simple CLI wrapper for the StarCoder code LLM using the Hugging Face inference api and wrapper library in node.js. This exposes the model's capabilities in a simple program inspired by old-school, text-based Unix tools and Github Copilot.

Features

• AI code generation at the command line!
• Saves your Hugging Face API access token locally in your home directory for subsequent runs
• Composable with other command-line tools using pipes, redirection, etc.
• it's free!

Installation

Requires:

• Node.js >= 18
• an account on the Hugging Face hub 🤗

Steps:

1. Accept the conditions for the bigcode/starcoder model and create a personal access token if you do not already have one.
2. Run npm install -g starcoder-cli to install the CLI globally
3. You will be prompted for your API access token the first time you run the command (this will be stored locally for subsequent runs).

Usage

~$ starcode --help
Usage: starcode [options] <code to complete>

Options:
      --version  Show version number                                   [boolean]
  -f, --file     Specify a file to get text input from                  [string]
      --help     Show help                                             [boolean]

For example:

~$ starcode "def print_hello_world():"
def print_hello_world():
    print("Hello World")

def print_hello_world_twice():
    print

Realistically, you'll probably want to use a source code file as the input instead of trying to enter large amounts of code as a command-line argument. For this, use the -f flag and provide a path to the existing (but incomplete) source code file. You can also use also redirect output to a file with shell redirection:

~$ starcode -f example.py > new_example.py

(this may fail with larger input files)

You can also use shell pipes with the starcode command. For example, this command reads the generated code aloud on macOS:

$ starcode -f new_example.py | say

Prompting Tips

• The StarCoder model used by this CLI is a text-completion model and is not tuned for instruction following or chat
• However, you may be able to use comments as a form of quasi-instructions that the model has exposure to from its training dataset
• Refer to the model paper (see "Further Reading" section) for the authors' dicussion of the model's limitations
• Code produced by this tool may have vulnerabilities or bugs

Contributing 🤝

Contributions are welcome and encouraged! Feel free to submit a PR if you have any improvements!

For now, I am using semistandard to check and fix style issues (this may change in the future).

Thank you in advance to anyone that wants to help with this project.

License

This project is licensed under the ISC license.

Acknowledgments 🙏

Thank you to BigCode project for the model that inspired this and Hugging Face

Further Reading

• StarCoder model paper
• HuggingFace.js Inference documentation notebook on Observable