Dataset Confidence

A training run can look good while the labels are noisy or the easy examples dominate, so the question is how much to trust the dataset. This page is dataset cartography: mapping and diagnosing datasets with training dynamics.

The same notes are in DISTILLING DATA.

The tool for that question is Dataset cartography mapping and diagnosing datasets with training dynamics - and its abstract starts from the same worry. Large datasets have become commonplace in NLP research. However, the increased emphasis on data quantity has made it challenging to assess the quality of data. We introduce Data Maps---a model-based tool to characterize and diagnose datasets. We leverage a largely ignored source of information: the behavior of the model on individual instances during training (training dynamics) for building data maps.

That behavior becomes something you can measure per row. This yields two intuitive measures for each example---the model's confidence in the true class, and the variability of this confidence across epochs---obtained in a single run of training. Experiments across four datasets show that these model-dependent measures reveal three distinct regions in the data map, each with pronounced characteristics.

Each region answers a different trust question. First, our data maps show the presence of "ambiguous" regions with respect to the model, which contribute the most towards out-of-distribution generalization. Second, the most populous regions in the data are "easy to learn" for the model, and play an important role in model optimization. Finally, data maps uncover a region with instances that the model finds "hard to learn"; these often correspond to labeling errors. Our results indicate that a shift in focus from quantity to quality of data could lead to robust models and improved out-of-distribution generalization.