Release AsyncWebRL model checkpoints on Hugging Face

Hi @JackBAI 🤗

I'm Niels and work as part of the open-source team at Hugging Face. I discovered your work through the Hugging Face Paper Pages and was impressed by the throughput speedups and SOTA results you achieved with AsyncWebRL!

The paper page lets people discuss your work and find related artifacts; you can also claim the paper as yours to show it on your public profile, and add GitHub/project page URLs.

It's great to see that you've already made the webgym_tasks dataset available on the 🤗 Hub! It'd be awesome to also host the trained model checkpoints (the Instruct and Thinking variants) on Hugging Face to improve their visibility and make them easily discoverable for the community.

Hosting on Hugging Face will allow people to find the models when filtering by tasks or tags, and we can link them directly to the paper page. If they are custom models, you can use the PyTorchModelHubMixin class to add from_pretrained and push_to_hub support, or simply use hf_hub_download for a one-liner download.

See here for a guide on uploading: https://huggingface.co/docs/hub/models-uploading.

We encourage pushing each checkpoint to a separate repository so that download stats and model cards work effectively.

Let me know if you're interested or need any help with the process!

Kind regards,

Niels
ML Engineer @ HF 🤗

Start with the Hugging Face model-uploading guide linked in the issue and inspect the repository for the trained Instruct and Thinking checkpoints alongside the already-published webgym_tasks dataset. Confirm how each checkpoint can be uploaded to its own repository with model-card and download support. Done means both variants are available on Hugging Face and can be linked to the paper page.