Hi @weigao266 馃
Niels here from the open-source team at Hugging Face. I discovered your work on Linear-MoE through our daily papers feature: https://huggingface.co/papers/2503.05447. The paper page allows for community discussion and artifact discovery. You can also claim authorship for enhanced visibility on your HF profile.
Your Linear-MoE system looks very promising! We noticed your GitHub repository (https://github.com/OpenSparseLLMs/Linear-MoE) contains training scripts and configurations for various model sizes and architectures, suggesting the existence of pretrained checkpoints. Would you be interested in releasing these checkpoints on the Hugging Face Hub? This would significantly boost their discoverability and allow the community to easily access and build upon your work.
We can add appropriate tags to the model cards to enhance searchability and link them directly to your paper page. If the models are PyTorch-based, the PyTorchModelHubMixin can simplify uploading with push_to_hub. Alternatively, hf_hub_download allows direct downloads from the hub.
We encourage pushing each checkpoint to a separate model repository for better tracking and download stats. Let me know if you're interested and if you need any assistance with the process!
Cheers,
Niels
ML Engineer @ HF 馃
Hi @weigao266 馃
Niels here from the open-source team at Hugging Face. I discovered your work on Linear-MoE through our daily papers feature: https://huggingface.co/papers/2503.05447. The paper page allows for community discussion and artifact discovery. You can also claim authorship for enhanced visibility on your HF profile.
Your Linear-MoE system looks very promising! We noticed your GitHub repository (https://github.com/OpenSparseLLMs/Linear-MoE) contains training scripts and configurations for various model sizes and architectures, suggesting the existence of pretrained checkpoints. Would you be interested in releasing these checkpoints on the Hugging Face Hub? This would significantly boost their discoverability and allow the community to easily access and build upon your work.
We can add appropriate tags to the model cards to enhance searchability and link them directly to your paper page. If the models are PyTorch-based, the
PyTorchModelHubMixincan simplify uploading withpush_to_hub. Alternatively,hf_hub_downloadallows direct downloads from the hub.We encourage pushing each checkpoint to a separate model repository for better tracking and download stats. Let me know if you're interested and if you need any assistance with the process!
Cheers,
Niels
ML Engineer @ HF 馃