Skip to main content
Analysis is the first step toward building a better model, but the real improvements come from what you do with your insights. After you’ve used the tools in a model run to identify your model’s strengths and weaknesses, the next step is to take targeted action. This is how you “close the loop” in your MLOps lifecycle: turning model-driven insights into higher-quality data and, ultimately, a more accurate model. Labelbox provides a suite of powerful actions that you can apply to any selection of data rows directly from your model run. This guide will walk you through each of these actions, explaining what they do, why they’re useful, and how to use them effectively.

​
Curate data for re-labeling

This is the most direct way to fix issues you discover during analysis. By sending data back to a labeling project, you can correct errors in your ground truth labels, add annotations to edge cases your model missed, or refine existing labels with greater precision. When to use this:
  • Correcting Labeling Errors: You discover that your model is being penalized for making a correct prediction because the ground truth label is wrong.
  • Addressing Model Failures: You identify a specific class of objects or a scenario (e.g., “blurry images,” “rainy conditions”) where your model consistently fails. Sending these examples for labeling helps your model learn from its mistakes.
  • Improving Annotation Quality: You notice that existing labels are not precise enough (e.g., bounding boxes are too loose), and you want your labeling team to refine them.

​
How to send data rows for re-labeling

  1. Select your data: In your model run, use filters to isolate the data you want to send for labeling. For example, filter for images where the model’s confidence was low but the IoU was also low (indicating a likely missed object). Select these data rows.
  2. Initiate action: Click the Manage selection button that appears at the bottom of the screen.
  3. Choose “Send to project”: From the action menu, select Send to project.
  4. Configure the batch: A dialog box will appear, prompting you to:
    • Select a project: Choose the labeling project where this data should be sent.
    • Set a batch name: Give the batch a descriptive name, like “Fix Bounding Boxes - Rainy Street Scenes”.
    • Set a priority: Assign a priority (Low, Medium, High) to the batch to signal its importance to your labeling team.
  5. Confirm: Click Confirm to create the batch and send the data to the selected project.
A new, prioritized batch is created in the queue of the target labeling project. Your labeling team can now access this batch and begin working on the data rows with the full context of your instructions, ensuring that the most critical labeling tasks are addressed first.

​
Augment your dataset with similar data

If you find one interesting or difficult example, there are likely many more like it in your datasets. The “Find similar” feature uses the power of vector search across your entire Catalog to instantly find more examples of the data you care about. This is a powerful form of AI-powered data curation that helps you scale your insights. When to use this:
  • Boosting Representation of Edge Cases: You find a few examples of a rare but important edge case (e.g., an occluded pedestrian). Use this action to find more examples to ensure your model can learn to handle this scenario.
  • Reducing Class Imbalance: Your model is underperforming on a minority class. Find more examples of that class from across your datasets to create a more balanced training set.
  • Active Learning Workflows: You want to find more data that is “hard” for your model. Find examples where your model was wrong, then use “Find similar” to surface a larger set of challenging data to label.

​
How to find similar data

  1. Select Representative Examples: In your Model Run, select one or more data rows that are good examples of the type of data you want to find.
  2. Initiate Action: Click Manage selection.
  3. Choose “Find similar in Catalog”: Select this option from the menu.
You will be taken to the Catalog, where Labelbox will perform an embedding-based similarity search. The results will be a collection of data rows from all your datasets that are visually or semantically similar to your selection. From here, you can review the results, select the most relevant examples, and send them to a project for labeling, creating a powerful, targeted dataset for your next model training run.

​
Create a dynamic slice

A Slice is a saved query that acts as a dynamic, “smart” folder for your data. Rather than creating a static list of data rows, a Slice continuously and automatically updates itself as new data is added or metadata changes. When to use this:
  • Tracking Key Segments: You want to constantly monitor your model’s performance on a critical subset of your data (e.g., “all images from camera-02” or “all data with a night-time metadata tag”).
  • Automating Data Curation: Create a Slice for “data with low confidence predictions” to automatically group your model’s most uncertain predictions for review.
  • Building Validation Sets: Create a Slice to represent a specific distribution of data that you want to use as a consistent validation set across multiple Model Runs.

​
How to create a slice

  1. Apply Your Filters: In your Model Run, apply the set of filters that defines the cohort of data you want to track.
  2. Save the Slice: Click the Save as Slice button at the top of the filter bar.
  3. Name Your Slice: Give your Slice a descriptive name that reflects the query, such as “Low Confidence Detections” or “Validation Set: Highway Scenes”.
The Slice is now saved and available in both the Model and Catalog tabs. You can select this Slice in any Model Run to instantly apply the saved filter set. The Slice will automatically include any new data that matches its criteria, providing a powerful way to automate your data management and analysis workflows.

​
Export data for offline analysis

While Labelbox provides a rich set of analysis tools, you may sometimes need to perform custom analysis or integrate with other tools in your MLOps stack. The export functionality allows you to download your data, predictions, and labels for use outside of Labelbox. When to use this:
  • Custom Reporting: You need to create custom visualizations or reports in a BI tool or a Python notebook.
  • Advanced Debugging: You want to perform deep, code-level debugging of your model’s predictions in your local development environment.
  • Sharing Results: You need to share a specific subset of data and predictions with colleagues who may not have access to Labelbox.

​
How to export data

  1. Select Your Data: Filter and select the data rows you wish to export.
  2. Initiate Action: Click Manage selection.
  3. Choose “Export data”: Select this option from the menu. You will be prompted to choose what to include in the export (e.g., data rows, ground truth, predictions, metadata).
  4. Confirm: Confirm the export. A job will be created to generate the export file.
You will be notified when your export is ready. You can then download a JSON or CSV file containing all the selected information, ready to be loaded into your custom analysis scripts or other tools.