2026/5/27 (Target)

May Update

This month's update centers on more flexible VQA evaluation and faster bulk editing in DataVerse. You can now use Free Text prompts in VQA prediction — for both single image testing and batch runs across a Data Slice — and drill into results with a redesigned VQA Metrics Report that filters by question and answer status. We've also added Batch Tag for editing tags across up to 1,000 images at once directly from Data Visualization, and Class-level Save to Ground Truth, which lets you update a single class instead of all annotations at once.

New Feature

Batch Tag

Batch edit tags across multiple images directly from the Data Visualization interface, eliminating the need to update tags one by one in the Annotation view — significantly improving efficiency for large-scale metadata management.

How to Use

1. Select the images to edit in Data Visualization (up to 1,000 at a time)

2. Click Edit Image Info → Edit Tags in the top-right corner

3. Toggle on the tags to edit, then add, remove, or clear as needed

4. Confirm the changes and wait for the success notification; results are available in Home → Notifications

VQA Tool

VQA add free text prompt input format and raw output

Use Free Text prompts in VQA prediction, both for single image testing and batch runs across a Data Slice. This removes the constraint of predefined question structures, making it faster to test different prompt formats, evaluate LLM compatibility, and compare accuracy across prompts using quantitative metrics.

Single Image Prediction

1. Go to Data Visualization → Review Image → Single Image Prediction

2. Switch Prompt Type to Free Text and enter your prompt in the textarea

3. Optionally configure Advanced Settings (Temperature, Max Tokens, Top P)

4. Click Run Prediction to send the prompt to the VLM

5. View results under Structured Result (mapped to ontology) or Raw Data (original VLM response)

Batch Prediction (Data Slice)

1. Select the Data Slice and click Prediction

2. Switch Prompt Type to Free Text and enter your prompt

3. Optionally configure Advanced Settings

4. Click Start Prediction — progress is tracked on the Prediction Status page

5. Once complete, results are available in Data Visualization; answers are mapped to ontology questions in order, with unmatched or missing answers marked as NA

VQA Metric Report - visualization

The VQA Metrics Report has been redesigned with a new evaluation tool that lets users drill into prediction results by question and answer status, making it faster to identify where prompts need improvement.

How to Use

1. Open Metrics Report and select a Data Slice

2. Use the Selector panel to filter images by question and answer status (Correct / Incorrect / NA)

3. Click into any image to review the Result tab — showing Ground Truth vs. Prediction per question — or switch to Raw Output for the full VLM response

4. Use Previous / Next to navigate across filtered images, or click View in Data Visualization to inspect the image in context

Metrics Report results are retained for 30 days. To keep a permanent record, download the results as a JSON file before they expire.

Annotation Enhancement - Class-level Save to Ground Truth

Save to Ground Truth now supports updating a specific class instead of applying changes to all annotations at once, significantly reducing review time when only a single class needs to be updated.

How to Use

1. Complete the Annotation Task review as usual

2. In Save to Ground Truth, select the target class to update

3. Choose the update mode (Overwrite / Append / Keep) — applied to the selected class only