Data Slice - Specific Subsets

Create endless possibilities for model training, validation, and testing with Data Slices.

Data Slice

Data Slice allows you to create countless datasets by filtering and combining data, which helps with model training, validation, and testing.

Hint

Use Data Visualization to explore your dataset and filter it down to create specific subsets, known as 'Data Slice'.

Quick Start

Data Slice. Click on the video below for a quick overview:

You can access all Data Slice lists on the Data Slice page and click "Details" to view each Data Slice's specifics.

Saving Data Slice

You can choose multiple datasets to combine into the data slice you need.

On the Data Visualization page, click "Save Data Slice" at the top to enter the data subset settings.

In the pop-up window, fill in the relevant information:

• Name: Name of the data slice

• Status: Increase the identification of the data subset to confirm whether it is unique and ready data.

• Select a method to save your data as a slice

  • Sampling Method: DataVerse offers a variety of sampling techniques to help users efficiently select data slices for analysis, training, or testing purposes. More about "Data Sampling"

  • Splitting Meothod: DataVerse offers four distinct options: Random, Sequence Independent, Class Balance and Tag Balance.More about "Data Splitting"

  • Image Deduplication: Automatically identify and remove visually similar images based on embedding proximity, and save the cleaned subset as a new data slice.

• Tag: Increase the identification of the data subset for easy searching.

• Description: Description

Once successfully created, you can view the created data subset on the Data Slice page.

About Data Slice page

• Start annotating the data by clicking "Start Annotation" and " + Annotation

• Make predictions by clicking "Prediction". + Prediction

• To export the data in VisionAI Format to a specified location, click "Export."

Furthermore, you can view the image status by clicking "View in Data Visualization."

VQA

• Prediction: Set up the details for annotation, and open the auto-labeling platform to operate labeling jobs.

• Run Metric: After run prediction results, use "Run Metrics" to compare them against your ground truth and evaluate VLM model performance. Review each question's accuracy rate to identify areas for improvement and guide your next iteration. Supported answer types: Boolean, Option, and Number.