Dataset Cartography: Mapping and Diagnosing Datasets with Training Dynamics

By Swabha Swayamdipta, Roy Schwartz, Nicholas Lourie, Yizhong Wang, Hannaneh Hajishirzi, Noah A. Smith, Yejin Choi. 2020

601 citations recorded by the source corpus.

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. 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. 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.

Subfields: Artificial Intelligence, Evaluation & Benchmarks

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Artificial Intelligence

In Evaluation & Benchmarks

Citations in the corpus

Understanding Dataset Difficulty with V-Usable Information (2021, positive): “We additionally consider slices in the dataset based on dataset cartography (Swayamdipta et al., 2020), which uses training dynamics to differentiate instances via their (1) confidence (i.e., mean probability of the correct label across epochs), and (2) variability (i.e., variance of the former).”
CommonsenseQA 2.0: Exposing the Limits of AI through Gamification (2021, background): “Indeed, recent research highlights the importance of diverse 27 and hard instances for robust out-of-domain generalization [16, 17].”
Dynabench: Rethinking Benchmarking in NLP (2021, neutral): “Recent work has also focused on evaluating model diagnostics through counterfactual augmentation (Kaushik et al., 2020), decision boundary analysis (Gardner et al., 2020; Swayamdipta et al., 2020), and behavioural testing (Ribeiro et al., 2020).”
GraDA: Graph Generative Data Augmentation for Commonsense Reasoning (2022, methodology): “Confidence and Variability.”
Challenges in Automated Debiasing for Toxic Language Detection (2021, background): “DataMaps (Swayamdipta et al., 2020) show the presence of distinct regions in a dataset— namely, easy, hard and ambiguous—defined with respect to a given model.”
Competency Problems: On Finding and Removing Artifacts in Language Data (2021, methodology): “Data ﬁltering A recent trend is to remove data from a training set that is biased in some way in order to get a model that generalizes better (Le Bras et al., 2020; Swayamdipta et al., 2020; Oren et al., 2020).”
Handling Bias in Toxic Speech Detection: A Survey (2022, methodology): “[120] investigated methods that have been efective on other NLU debiasing tasks [54, 99].”
A Long Way to Go: Investigating Length Correlations in RLHF (2023, methodology): “Results First, we note that when examining “cartography” plots (Swayamdipta et al., 2020) examining the mean ( c i ) and variance ( σ ( c i ) ) of different c i (see Appendix B.1), we find that the values are largely centered at zero, suggesting that reward models are not able to make progress on…”
Evaluation Examples are not Equally Informative: How should that change NLP Leaderboards? (2021, background): “For example, Swayamdipta et al. (2020) and Pleiss et al. (2020) (Boyd-Graber and Börschinger, 2020), prioritizing examples most likely to reveal interesting distinctions between submitted systems.”
Metadata Archaeology: Unearthing Data Subsets by Leveraging Training Dynamics (2022, methodology): “No learning rate decay is applied in this case.”
Identifying Mislabeled Data using the Area Under the Margin Ranking (2020, background): “Many approaches use training loss as a proxy for label quality [19, 52, 54, 60, 62, 63].”
DataComp: In search of the next generation of multimodal datasets (2023, methodology): “[129] Swabha Swayamdipta, Roy Schwartz, Nicholas Lourie, Yizhong Wang, Hannaneh Hajishirzi, Noah A.”
Interpretable deep learning: interpretation, interpretability, trustworthiness, and beyond (2021, methodology): “We hope the taxonomy can shed light on future improvements/extensions on explaining Table 2 List of interpretation algorithm publications Methods Publications (non-exhaustive) LIME and variants LIME [137], Anchors [138], SHAP [110], RISE [127], MAPLE [130] Global interpretation LIME-SP [137], NormLI”
SOTOPIA: Interactive Evaluation for Social Intelligence in Language Agents (2023, background): “To fully see the difference between humans and models, we select the most challenging scenarios following Dennis et al. (2020); Swayamdipta et al. (2020).”
CREAK: A Dataset for Commonsense Reasoning over Entity Knowledge (2021, methodology): “We carefully examine such artifacts in our datasets using quantitative tools [Swayamdipta et al., 2020, Gardner et al., 2020] as well as qualitative inspection.”
GraDA: Graph Generative Data Augmentation for Commonsense Reasoning (2022, methodology): “Swayamdipta et al. (2020) propose the model confidence (µ̂i) and variability (σ̂i) measures to identify the effect of data samples on the model’s generalization error.”