Changelog

New features

New Insert Plugins menu and Testing interface

There are a number of new features and changes related to plugins:

• We have renamed "Custom Scripts" to "Plugins." This is reflected in the UI and in our docs.
• There is a new Insert Plugins menu available. From here you can insert a pre-built plugin that you can customize as necessary.
• When you add a plugin, you will see a new Testing panel below the plugin editing field. You can use this to verify what events are triggered, manually trigger events, and modify the sample data as necessary.
• To accompany the new Insert Plugins menu, there is a new Plugins gallery in the documentation that discusses each option and has information on creating your own custom plugs.
• There is also a new setting that allows you to restrict access to the Plugins tab to Administrator users. By default, it is also available to Managers. If you would like this set, contact your account manager or open a support ticket.

New Label Studio Home page

When you open Label Studio, you will see a new Home page. Here you can find links to your most recent projects, shortcuts to common actions, and links to frequently used resources.

Support for Google Cloud Storage Workload Identity Federation (WIF)

When adding project storage, you now have the option to choose Google Cloud Storage WIF.

Unlike the the standard GCS connection using application credentials, this allows Label Studio to request temporary credentials when connecting to your storage.

For more information, see Google Cloud Storage with Workload Identity Federation (WIF).

Feature updates

Prompts model support changes

We have added support for the following Gemini models:

• gemini-2.0-flash
• gemini-2.0-flash-lite

We have removed support for the following OpenAI models:

• All gpt-3.5 models

Improved tooltips

• Improved tooltips related to pausing annotators.
• Added a tooltip to inform users that Sandbox projects are not supported in Prompts.

Miscellaneous

• Ensured that when a user is deactivated, they are also automatically logged out. Previously they lost all access, but were not automatically logged out of active sessions.
• Multiple performance improvements for our AI Assistant.

Breaking changes

Label Studio is transitioning from user to roles in AWS S3 IAM.

This affects users who are using the Amazon S3 (IAM role access) storage type in projects.

Before you can add new projects with the Amazon S3 (IAM role access) storage type, you must update your AWS IAM policy to add the following principal:

arn:aws:iam::490065312183:role/label-studio-app-production

Please also keep the the previous Label Studio principal in place to ensure that any project connections you set up prior to April 7, 2025 will continue to have access to AWS:

arn:aws:iam::490065312183:user/rw_bucket

For more information, see Set up an IAM role in AWS.

Security

• Made security improvements around the verbosity of certain API calls.
• Made security improvements around SAML.
• Made security improvements around project parameter validation.
• Made security improvements around exception error messages.
• Upgraded Babel to address vulnerabilities.

Bug fixes

• Fixed an issue where users were unable to edit custom agreement metrics if using manual distribution mode.
• Fixed an issue where role changes with LDAP were broken.
• Fixed error handling for Prompts refinement.
• Fixed an issue where Exact frames matching for video was always showing as an option for agreement metrics regardless of whether the labeling config referenced a video.
• Fixed an issue where the prediction-changed value was not being reset after making manual changes to pre-annotations.
• Fixed an issue where a "Script running successfully" message continuously appeared for users who had plugins enabled.
• Fixed an issue where interacting with the Manage Members modal would sometimes throw an error.

New features

Prompts is now generally available

Prompts is now available in Label Studio Enterprise SaaS.

Prompts is a tool that leverages LLMs to automatically generate predictions for your labeling tasks, allowing you to pre-annotate data and quickly bootstrap projects. It can also be used to evaluate your LLM prompts against ground truth data.

For more information, see the Prompts documentation and Accelerate Labeling and Model Evaluation with Prompts—Now Generally Available.

Pause an annotator

There is a new action to pause annotators. This is available from the Members dashboard and via the API.

For more information, see Pause an annotator.

Personal access tokens

There is a new type of token available for API access. The new tokens use JWT standards and, unlike the current tokens in use, allow you to set an expiration date.

You can enable or disable these tokens from the Organization page. Once enabled, they will be available for users to generate from their Account & Settings page. Legacy tokens can still be used unless disabled from the organization level.

For more information, see Access tokens.

Deep link annotations and regions

You can now link directly to specific annotations or regions within an annotation. These actions are available from the labeling interface in the overflow menus for the annotation and the region.

Feature updates

Set ground truths by user

There is a new action from the Data Manager that allows you to mark the annotations submitted by a specific user as ground truth annotations.

Known issues

Users in the Manager, Annotator, and Reviewer role cannot currently retrieve their API access tokens from the Label Studio UI.

This does not affect the functionality of any existing scripts or automations currently using their access tokens.

Bug fixes

• Fixed an issue that was causing intermittent loading errors in the Data Manager.
• Fixed an issue where filters were not being respected when performing bulk annotation actions.
• Fixed an issue where bulk annotation was sometimes failing with a 500 error.
• Fixed an issue where bulk annotation was not respecting the Allow empty annotations setting.
• Fixed an issue that caused incorrect video frame to be rendered when pausing.
• Fixed an issue where users were getting their role reset if they were provisioned via SCIM without an assigned group.
• Fixed an issue where an error was sometimes thrown when loading the workspaces list.
• Fixed an issue that would cause an API error when switching workspaces and fetching the incorrect page of projects.

New features

Introducing the Prompts tool

The Prompts tool leverages ChatGPT to help you evaluate and refine your LLM prompts. You can also use Prompts to generate predictions for automating your labeling process, and to quickly bootstrap labeling projects.

For more information, see Automate Data Labeling with HumanSignal and our Prompts documentation.

Refreshed Label Studio interface

The Label Studio UI has been upgraded with updated colors and fonts, giving it a sleek new look while maintaining the same intuitive navigation you're familiar with. All Label Studio tools, features, and settings are still in the same place, ensuring a smooth transition.

Feature updates

Improved performance on the Projects list page due to improvement on the API level.

Breaking changes

Fixed an issue with Google Cloud Storage when the connection has the Use pre-signed URLs option disabled. In these situations, Google was sending pre-signed URLs with the format https://storage.googleapis.com rather than sending BLOBs.

With this fix, Google Cloud Storage will begin returning BLOBs/base64 encoded data when Use pre-signed URLs is off. This means that Label Studio will start reading data from Google Cloud Storage buckets, which can result in large amounts of data being sent to your Label Studio instance - potentially affecting performance.

Bug fixes

• Fixed an issue where Redis storage connections were causing errors due to a missing field in the storage form (Storage Title).
• Fixed an issue where annotators were not being prompted to leave a comment when skipping a task, even though the project settings required them to do so.
• Fixed an issue where connected ML backends were unable to return more than one prediction per task.