Chapter 8. Dataset Engineering
The quality of a model depends on the quality of its training data. The best ML team in the world with infinite compute can’t help you finetune a good model if you don’t have data. The goal of dataset engineering is to create a dataset that allows you to train the best model, ideally within your allocated budget.
As fewer companies can afford to develop models from scratch, more are turning to data to differentiate their AI performance. As models demand more data, data handling becomes more challenging and demands more investments in talent and infrastructure.1
Data operations have evolved from side tasks that people handle when they have time to dedicated roles. Many AI companies now employ data labelers, dataset creators, and data quality engineers, either integrated into or working alongside their core engineering teams.
If the model landscape is confusing enough with numerous offerings, the data landscape is even more complex, with an ever-growing array of datasets and techniques being introduced. This chapter gives you an overview of the data landscape and considerations to take into account when building your own dataset.
It begins with data curation, addressing questions like What data do you need? How much? What does it mean for data to be of high quality? It then discusses techniques for data synthesis and processing. Data curation, generation, and processing don’t follow a linear path. You’ll likely have to go back and forth between different ...
Get AI Engineering now with the O’Reilly learning platform.
O’Reilly members experience books, live events, courses curated by job role, and more from O’Reilly and nearly 200 top publishers.
