Overview of Manifest Management

When using ModelArts, perform operations such as labeling data, training a model, performing inference, managing datasets, and publishing a model in AI Gallery. All these operations are based on datasets. To standardize the use of datasets in various application scenarios and ensure the flexibility of dataset management, this document describes the manifest file with dataset management APIs and specifications included.

• A manifest file defines the mapping between labeled objects and content. The manifest file can contain only unlabeled data, for example, a created dataset that has not been labeled.
• A manifest file is encoded using UTF-8. Therefore, the programs processing manifest must support UTF-8.
• In a manifest file, the text classification source allows non-English characters.
• You can create a manifest file or obtain such a file using a third-party tool or ModelArts.
• Any valid file name is allowed for a manifest file.