Work Flow | Graviton

First, you can familiarise yourself with Graviton by checking out these which guide you through the basic Graviton functionalities, including account registration, team creation, and dataset uploading. For advanced administration management, check out .

Upload & Manage: Congratulations! You have successfully uploaded your first dataset to Graviton. Owing to our data-centric design, your dataset is the main component throughout the entire workflow, getting improved over time by our data curation and data mining tools. As such, you would want your data to be reasonably organised. Here, we provide tools to upload, inspect and version-control your datasets.

Curate: It is often the case that you have collected an enormous amount of unlabeled data, but only want to work with a subset of it. Understand this, we provide the deduplication tool that removes duplicate data from your dataset, effectively saving storage cost and running time of subsequent tools. We also provide active learning, a powerful, automated tool that helps to select the best subset of your data for annotation.

Annotate: We provide machine-assisted annotations tools and on-demand annotation services to help you obtain high-quality ground truths, for which you can control the data quality using our QC system. After annotating, you can to a wide range of common formats as well as gain meaningful insights from the dataset using our dashboard.

Evaluate: After you finish training your ML models from the high-quality dataset obtained from the previous step, it is time to evaluate your model performance. We provide Model Evaluation, an automatic, visual, and qualitative way of interpreting your model behaviors.

Find Edge Cases: From Model Evaluation, you will be able to manually inspect your model predictions and find systematic errors of your model. We provide Edge Case Mining to help you find objects in your large unlabeled dataset that are similar to your identified edge cases.

Select Core Set: helps in finding a subset of data points that maintain a similar level of quality as the full dataset.