Release MTDEval, P^2-MTD, Daily-MTD On Hugging Face

Hey guys! 👋 Niels from the Hugging Face open-source team reached out about our paper, and we're super stoked to share the deets on how we're making our MTDEval model checkpoints, P^2-MTD, and Daily-MTD datasets available on the Hugging Face Hub. This is a game-changer for improving visibility and accessibility, so let's dive in!

Why Hugging Face?

Hugging Face is the go-to platform for the AI community, and having our resources there means more researchers and developers can easily find and use our work. It's like setting up shop in the heart of the AI universe! 🌌

Niels stumbled upon our work through Hugging Face's daily papers feature, which is pretty cool. Our paper got featured here, and this kind of visibility is invaluable. Hugging Face also lets people discuss papers and find related artifacts like models, datasets, and demos. You can even claim your paper, which adds it to your public profile and allows you to link GitHub and project pages. This is all about making our work more discoverable and connected.

Our main goal is to make our MTDEval model checkpoints and the P^2-MTD and Daily-MTD datasets readily available on the Hugging Face Hub. This will significantly improve their discoverability, allowing more people to utilize our resources. Hugging Face's tagging system will enable users to easily filter and find our contributions within their extensive collection of models and datasets. This increased accessibility is crucial for fostering collaboration and accelerating research progress in our field.

Boosting Discoverability

One of the biggest advantages of using Hugging Face is its discoverability features. By adding tags, our datasets and models will pop up when people filter through models and datasets. It's like having a spotlight on our work! ✨

Collaboration and Community

Hugging Face is more than just a platform; it's a community. By making our resources available here, we're inviting collaboration and feedback, which is essential for refining and improving our work.

Uploading Models: A Step-by-Step Guide

So, how do we actually get our models onto Hugging Face? Niels pointed us to this handy guide. The process is straightforward, and we can leverage some cool tools to make it even easier.

PyTorchModelHubMixin

This is a nifty class that adds from_pretrained and push_to_hub methods to any custom nn.Module. Basically, it simplifies the process of loading and pushing models. It's like having a turbo button for model sharing! 🚀

hf_hub_download

For a more streamlined approach, we can use the hf_hub_download one-liner to grab checkpoints directly from the Hub. It’s super convenient for quick access and testing. Think of it as a direct download link for your models.

Best Practices: Separate Repos for Checkpoints

Niels suggested pushing each model checkpoint to its own repository. This might seem like extra work, but it's worth it. It allows for accurate download stats and makes it easier to track the performance of different checkpoints. Plus, we can link each checkpoint to the paper page, creating a comprehensive resource hub. It’s like organizing a well-stocked library for our models!

Uploading models to Hugging Face involves a few key steps to ensure they are easily accessible and well-documented. First, we need to leverage the PyTorchModelHubMixin class, which simplifies the process by adding from_pretrained and push_to_hub methods to our models. This allows us to seamlessly load and upload models with just a few lines of code. Alternatively, the hf_hub_download one-liner provides a quick way to download checkpoints directly from the Hub, making it easier to test and integrate models into existing workflows. Following the best practice of pushing each model checkpoint to a separate repository is crucial for maintaining organization and facilitating performance tracking. This approach allows for accurate download statistics and enables us to link each checkpoint directly to our paper page, creating a comprehensive resource hub for our research community.

By adopting these methods, we ensure our models are not only accessible but also easy to use and track, fostering broader adoption and collaboration within the AI research community. This strategic approach to model sharing significantly enhances the impact and reach of our work.

Uploading Datasets: Making Data Accessible

Datasets are the lifeblood of AI research, so making ours accessible is a top priority. Niels gave us a guide for this too, and it looks super straightforward.

The Magic of load_dataset

Imagine being able to load a dataset with just one line of code. That's the power of Hugging Face's load_dataset function! It's like having a universal key to unlock any dataset on the Hub.

from datasets import load_dataset

dataset = load_dataset(