Labelbox lets you slice and dice your model predictions, annotations, and metrics. Powerful filtering and sorting make it simpler and more efficient to:
  • Analyze the performance of neural networks
  • Find low-performing slices of data
  • Surface labeling mistakes
  • Identify high-impact data to label in order to model performance
Model runs contain versioned snapshots of annotations, predictions, and metadata. These versions persist over time, which means that you can search model runs for values even if they’re changed elsewhere in the Labelbox platform. Filters let you search model runs for specific values.

​
Supported attributes for search and filter

Below are the attributes you can search and filter by in the Model product.
AttributeDescriptionExamples
AnnotationFind data rows with labels that contain or do not have certain counts of annotationsShow image assets where X annotation was used at least N times.
PredictionPredictions uploaded to the model runShow text assets where the model predicted X at least N times.
Data rowFilter on any data row identifiers (e.g., global key, data row ID, and external ID)Show all image assets where the global key does not contain X.
LabelFind data rows by the ID of associated labels.Show all text assets where a label ID contains X.
DatasetFind data rows that belong to a particular dataset.Show all image assets uploaded to dataset X.
ProjectFind data rows that are associated with a specific labeling projectShow all text assets uploaded to project X.
MetricsModel run metrics (auto-generated and custom)Show image assets with specific values for precision, recall, intersection over union, etc. Filter on aggregated or per prediction metrics
MetadataFind data rows that contain a certain metadata field and/or valueShow text assets that were captured between two datetimes

​
Filters

You can think of filters as pyramids with layers of logical sequence. Each filter condition is a layer that limits the data rows in the view. You can think of each layer as an “AND” condition in a logical construct. Within a layer, you can add additional match conditions (“OR” conditions). The final results count reflects the data rows that match all currently specified conditions. The following screenshot shows several AND conditions, including:
  • Metrics where confidence is between 0.538 and 0.84.
  • Metrics where IoU is between 0.01 and 0.545.
  • Data rows with an annotation of kite or a prediction of person.
In this example, 46 data rows match all conditions.

This filter combines multiple different conditions.

​
Filter on custom metrics per prediction basis

You can upload custom metrics to model runs and then them in model view filters. For help uploading custom metrics, see Model Diagnostics - Custom Metrics demo. You can filter predictions by auto-generated metrics, by custom metrics, or both. This lets you create a unique view, such as filtering by a specific custom metric and then refining that with values in specific ranges, as shown in the following image. You can also combine such filters with other filters to create even more tailored views. To do this, select Metrics as your filter type and then select per prediction. Next, select the specific feature you want to use to filter your model view. Next, add additional filter conditions to create the specific view you need.

Use the "Metrics" filter, then select "per prediction".

​
Sort by

You can sort data on any combination of auto-generated metrics and/or custom metrics. To do so, click on Sort by, then pick the desired metric and decide if the sorting should be in ascending order (Asc) or descending order (Desc). You can pick metric filtering per prediction for sorting a particular metric or aggregated for sorting by the arithmetic mean of the metric.

Sort by option in Model Run tab