**Croissant: A Metadata Format for ML-Ready Datasets** This paper introduces Croissant, a metadata format designed to enhance the discoverability, portability, reproducibility, and interoperability of machine learning (ML) datasets. Croissant simplifies the use of datasets by ML tools and frameworks, making them more accessible and reusable. The format is supported by several popular dataset repositories, including Hugging Face, Kaggle, and OpenML, and can be loaded into various ML frameworks. **Key Features:** - **Dataset Metadata Layer:** Contains general information such as name, description, and license. - **Resources Layer:** Describes the source data, including individual files and file sets. - **Structure Layer:** Organizes the structure of resources using RecordSets, which can handle various data types and support data manipulation. - **Semantic Layer:** Provides ML-specific interpretations, including custom data types and dataset organization methods. **Integration and Support:** - **Data Repositories:** Croissant has been integrated into Hugging Face, Kaggle, and OpenML, with over 400,000 datasets in the format. - **Data Loaders:** A standalone Python library supports validation, creation, and serialization of Croissant dataset descriptions. - **Croissant Editor:** A tool for visually creating and modifying Croissant datasets, integrating with the Responsible AI extension. - **Dataset Search:** Google Dataset Search supports Croissant datasets, allowing users to search for them across repositories. **Future Work:** - **Community Engagement:** Encouraging dataset repositories and tool developers to adopt Croissant. - **Semantic Development:** Guiding further development of ML-specific aspects based on user feedback. - **Interoperability:** Potential adoption by other fields to increase interoperability in data repositories and processing frameworks. **Conclusion:** Croissant aims to evolve based on user feedback and emerging needs in the field of machine learning, promoting responsible AI practices and enhancing the overall efficiency of ML data management.**Croissant: A Metadata Format for ML-Ready Datasets** This paper introduces Croissant, a metadata format designed to enhance the discoverability, portability, reproducibility, and interoperability of machine learning (ML) datasets. Croissant simplifies the use of datasets by ML tools and frameworks, making them more accessible and reusable. The format is supported by several popular dataset repositories, including Hugging Face, Kaggle, and OpenML, and can be loaded into various ML frameworks. **Key Features:** - **Dataset Metadata Layer:** Contains general information such as name, description, and license. - **Resources Layer:** Describes the source data, including individual files and file sets. - **Structure Layer:** Organizes the structure of resources using RecordSets, which can handle various data types and support data manipulation. - **Semantic Layer:** Provides ML-specific interpretations, including custom data types and dataset organization methods. **Integration and Support:** - **Data Repositories:** Croissant has been integrated into Hugging Face, Kaggle, and OpenML, with over 400,000 datasets in the format. - **Data Loaders:** A standalone Python library supports validation, creation, and serialization of Croissant dataset descriptions. - **Croissant Editor:** A tool for visually creating and modifying Croissant datasets, integrating with the Responsible AI extension. - **Dataset Search:** Google Dataset Search supports Croissant datasets, allowing users to search for them across repositories. **Future Work:** - **Community Engagement:** Encouraging dataset repositories and tool developers to adopt Croissant. - **Semantic Development:** Guiding further development of ML-specific aspects based on user feedback. - **Interoperability:** Potential adoption by other fields to increase interoperability in data repositories and processing frameworks. **Conclusion:** Croissant aims to evolve based on user feedback and emerging needs in the field of machine learning, promoting responsible AI practices and enhancing the overall efficiency of ML data management.