Datasets:

---

jhu-clsp

/

mmBERT-midtraining-data

like 1

Follow

Center for Language and Speech Processing @ JHU 213

This PR significantly improves the dataset card for mmBERT Mid-training Data by:

• Updating the task_categories metadata from fill-mask to feature-extraction. While the dataset is used for masked language model pre-training, the resulting mmBERT models are primarily used for downstream tasks like classification, retrieval, and generating embeddings, which fall under feature-extraction.
• Adding long-context and masked-language-modeling to the tags metadata to better reflect the dataset's use in training models with extended context capabilities and the specific pre-training objective.
• Renaming the existing "Usage" section to "Dataset Usage" for better clarity, distinguishing it from how the models trained with this data are used.
• Adding a comprehensive "Model Usage Examples" section with code snippets directly from the mmBERT GitHub repository. This new section demonstrates how to use the mmBERT models (trained using this dataset) for tasks such as obtaining multilingual embeddings, masked language modeling, classification, multilingual retrieval, and long context processing, enhancing the discoverability and utility of the artifact.

All existing links, including the Arxiv paper link, remain unchanged as per the guidelines.