Dataset planning
Developing an AI system to solve a use case requires developing two logically distinct solution stacks: the AI evaluation stack and the AI system stack.
The AI evaluation stack consists of a suite of evaluation datasets, associated mechanisms to produce the datasets, and the set of release criteria and metrics used to evaluate the system on the data. An evaluation dataset is used to test the performance of the AI system against one or more release criteria. An evaluation dataset may consist of inputs and their expected outputs, or just inputs where a separate mechanism (for example, a human workforce or another AI model) is used as part of the evaluation stack to evaluate the correctness of an AI system output for the input in the evaluation dataset.
The AI system stack consists of the AI system itself, required training or auxiliary datasets, and mechanisms needed to produce them. Training datasets are used for building the AI system and support learning, validation, calibration, and component and parameter selection. Auxiliary datasets provide information required during operation (for example, the policies used in automated ground truth checks or the document libraries used in a retrieval-augmented generation (RAG) system).
You can develop AI evaluation and AI system stacks independently. For example, your team and your downstream stakeholders may have their own AI evaluation stacks. Your downstream stakeholder may build their stack based on your documentation (see focus area User Guidance) and still end up making different tactical decisions about the design of their evaluation datasets. This will yield different evaluations of AI system performance.
Alternatively, you may discover that your ideal robustness evaluation dataset has a large statistical power requirement, so you build it out in increments, which results in your robustness performance dropping as each increment is added, even though your AI system stack remains unchanged.
The dataset planning focus area provides best practices for designing the datasets needed in both stacks. The System Planning focus area covers designing the AI system and the Evaluate and Release focus area covers using the evaluation stack to assess the AI system.
Datasets are often thought of as a fixed set of data objects (inputs or input and output pairs, in the case of training or evaluation datasets) that reside in permanent storage. However, the datasets used in training, evaluation, and operation are often dynamically constructed from the static stored version. For example, given a use case of recognizing cats and dogs in images and a static dataset of images labeled cat or not cat, a builder might construct an evaluation dataset for robustness by augmenting each image in the static dataset with rotated versions of the image, increasing the sensitivity of the test datasets to a key confounding variation (image rotation) without needing new labels. The best practices in the Dataset Planning focus area are ideally applied to the final datasets, including dynamic augmentation. However, but this may not be feasible as some static datasets may be augmented dynamically online throughout training.
