The debugging workspace
for machine learning
Startups and Fortune 500s alike use Unbox to track and version models, uncover errors, and make informed decisions on data collection and model re-training.
testing & debugging
Uncover & explain errors
Filter through your model’s predictions to quickly identify common failure patterns and tricky edge-cases alike. Our explainability techniques and automatic error clustering mechanisms help you take corrective action before you ship.
Systematically boost performance
Take corrective action to improve models through clearly defined suggestions. Retrain your model with synthetic data to make it more robust to discovered failure patterns, identify and relabel mislabeled data in bulk, alter model architectures or pre-processing schemes, and more. Finally, a repeatable recipe to rapidly build production-ready models.
Use synthetic data to make your model more robust to failure patterns. E.g. an error class that is very common, a feature range that is underrepresented, or a class with low accuracy.
Data cleaning & relabeling
Fix deficiencies in your data by re-labeling mislabeled data, using class-imbalance to adjust your dataset composition, or prioritizing what next batch of data needs to be collected or annotated.
Tune model architecture
Use insights to alter your model architecture, hyperparameters or preprocessing scheme. E.g. change your tokenizer if the model doesn’t recognize "happpyy" as positive.