The Data-Centric AI Playbook

An interactive guide to engineering high-quality datasets when real-world data is scarce, noisy, or insufficient.

The Paradigm Shift to Data-Centric AI

For decades, AI research focused on building better models. The new frontier of progress, however, lies in systematically engineering the data itself. This shift is driven by critical bottlenecks in the model-centric approach.

0

% of AI Projects Fail

due to issues with data quality, not model flaws.

0

% of Research was Model-Centric

historically, leading to a focus on code over data.

2023

Projected Year of Data Exhaustion

for high-quality public text data, forcing a move to data engineering.

Toolkit: Programmatic Labeling with Weak Supervision

When you have abundant unlabeled data but no labels, weak supervision allows you to programmatically create a large training set by encoding domain knowledge into heuristic "Labeling Functions" (LFs).

The Weak Supervision Pipeline

1

Write Labeling Functions (LFs)

Encode heuristics (e.g., keyword searches, patterns, LLM prompts) to programmatically vote on labels or abstain.

2

Train Generative Label Model

This model learns the accuracies and correlations of your LFs by observing their agreements and disagreements—no ground truth needed.

3

Generate Probabilistic Labels

The output is a full training set with "soft" labels (e.g., 90% Class A), capturing the model's aggregate confidence.

4

Train Discriminative End Model

A powerful end model (e.g., a Transformer) learns from the probabilistic labels to generalize beyond the simple heuristics of the LFs.

Toolkit: Efficient Data Selection with Active Learning

When your labeling budget is limited, Active Learning (AL) helps you maximize model performance by intelligently selecting the most valuable data points for manual annotation.

The Query Strategy Explorer

The "brain" of an active learner is its query strategy. Select a strategy below to understand its core principle for choosing which data to label next.

Toolkit: Creating New Data with Generative Methods

When you need to fill gaps in your dataset, cover rare edge cases, or simply create more data, generative methods provide the solution. Optimal Transport offers a principled, geometric framework for this task.

Principled Augmentation with Optimal Transport

Naive Interpolation (e.g., Mixup)

Simply averaging data points can create unrealistic samples that fall "off" the true data manifold, harming model training.

Wasserstein Barycenters (OT)

OT finds a geometric "average" that respects the data's structure, producing realistic, high-fidelity samples.

Build Your Strategy: A Unified Framework

These techniques are most powerful when combined. Use this interactive guide to determine the best data-building strategy for your specific problem.

Recommended Strategy:

Your recommended workflow will appear here...