A reading list for when your data is the problem, not the model

What it does data-centric-AI is a curated bibliography that organizes papers, code repositories, and tutorials around the premise that systematically engineering data matters as much as choosing a model architecture. The maintainers structure resources into a three-part framework—training data development, inference data development, and data maintenance—and openly admit the list is incomplete and selective, favoring papers that introduce distinct ideas over exhaustive coverage.

The interesting bit The list doubles as a field guide and a gentle manifesto. The authors draw a sharp, useful line between “data-centric” (engineering the data itself) and “data-driven” (using data to guide model development), and they use GPT and Segment Anything as case studies for why data work dominates model tweaks.

Key highlights

• Taxonomy spans the full lifecycle: data collection, labeling, preparation, augmentation, prompt engineering, and ongoing maintenance.
• Mixes canonical tools like Snorkel and Aurum with recent work on active learning, weak supervision, and foundation-model data curation.
• Bundles the authors’ own survey papers, a KDD 2023 tutorial, and blog posts that map the framework onto GPT and SAM.
• Hosts community channels (Slack, QQ, WeChat) for a topic that usually lacks a centralized water cooler.
• Explicitly warns that it is “unfeasible to encompass every paper,” so expectations are set honestly.

Caveats

• It is a reading list, not a toolkit; most entries are paper links, so practitioners hunting for off-the-shelf libraries will need to follow trails.
• Coverage is intentionally spotty in places, and the README notes that it selectively chooses papers rather than aiming for completeness.

Verdict Bookmark it if you are a researcher or MLOps engineer building a case for data-quality investment, or if you need a structured on-ramp to the field. Look elsewhere if you want a single executable framework or a hands-on coding tutorial.