Contributing to the LM Zoo

Anyone can contribute models to the LM Zoo. This page provides instructions on creating a LM Zoo-friendly Docker image, and on submitting your image for Zoo review.

Language models in the LM Zoo are Docker images which expose a set of binary programs conforming to the Language model Docker API. At a high level, the steps you’ll need to take to prepare such an image are

  1. Acquire open source code and model checkpoints for the language model you’d like to contribute.

  2. Prepare a Docker image containing the code and model checkpoint, and verify that you can successfully run the model from within the container. Write binary scripts conforming to the Language model Docker API which invoke your model.

  3. Submit your language model as a pull request to the cpllab/lm-zoo Github repository.

Each of these steps are described in further detail below.

Acquiring open-source code

Because the LM Zoo system runs on Docker, you are welcome to contribute code which uses any programming language and framework you favor, so long as you can provide a supporting Docker image.

You are also welcome to contribute language model images for models which are not your own, so long as this complies with the license of the model developers. When in doubt, ask the developers for permission first!

Preparing a Docker image

If you’re not familiar with building your own Docker images, we recommend that you first read the official Docker documentation.

  1. Create your own fork of the lm-zoo repository.

  2. Create your own subdirectory in the models directory. We recommend copying from the template directory for starters. Modify the Dockerfile and add wrapper scripts as necessary until your model conforms to the Language model Docker API.

  3. You can build and test your wrapper as necessary using the file scripts/build_and_test.sh:

    ./scripts/build_and_test.sh <my_directory_name> <docker_target>
./scripts/build_and_test.sh _template myname/mylm

A few tips on building an image:

Docker build context

All LM Zoo Docker images are built with the root directory of the LM Zoo as build context. This allows each Docker build routine to copy in shared scripts, for example, from the shared directory. Structure your build files accordingly – see the template directory’s Dockerfile and its use of a build argument MODEL_ROOT as an example.

Shared scripts

The shared directory provides several useful shared binary scripts (e.g. unkify and spec) which you may be able to reuse for your image. We offer these scripts to avoid forcing contributors to reinvent mundane code for every image they build.

Submitting your language model

Once you’ve confirmed that all tests pass for your image, create a pull request at the lm-zoo Github repository. Your pull request should follow this template (this should be automatically inserted when you create a pull request).

The LM Zoo maintainers will evaluate your submission and respond on the pull request thread. Thanks for contributing!