Here’s what’s new:
🏷️ The Auto Labeling Panel: Interactive human-in-the-loop annotation powered by foundation models
For many teams, annotation has often been a fragmented process involving disparate tools that make it difficult to apply labels that (1) are correct, and (2) actually lead to improved downstream model performance.
In FiftyOne Enterprise 2.14.0, we have unified the curation + annotation + evaluation workflow. The new Auto-Labeling Panel allows you to configure, execute, and review model predictions entirely within the FiftyOne App.
- Configure in the UI: Select your model directly from the panel. Choose among SOTA zero-shot models or fixed-vocabulary models, as well as a t-shirt size that balances speed and performance. Define your classes of interest, confidence thresholds, and applied label fields without writing a single line of code.
- Generate at scale: Once configured, click "Auto Label." The job is offloaded to your compute cluster, processing thousands of images in parallel. You can continue exploring your dataset while the job runs in the background.
- Interactive review: As predictions flow back into the dataset, the panel lets you interactively review them. You can:
- Filter predictions by confidence score.
- Bulk-approve high-confidence labels.
- Isolate low-confidence "edge cases" for manual refinement.
- Discard incorrect labels instantly.
This creates a tight feedback loop, and allows both annotators and engineers to focus their efforts on the edge cases and critical scenarios that truly move the needle.
🚀 Native Kubernetes Orchestrator: Ephemeral compute at scale
To support compute-intensive tasks at scale, your infrastructure needs to be elastic. FiftyOne Enterprise 2.14.0 introduces the Built-in Kubernetes Orchestrator. This new backend integrates directly with the Kubernetes API to treat compute as a utility instead of a static resource.
- True ephemerality: When a user triggers a task like an embedding computation or auto-labeling job, the system spins up a Kubernetes Job specifically for that task. When the task finishes, the pod is terminated. No more paying for idle GPUs.
- Hardware efficiency: Support for scheduling mechanisms like Node Affinity and Tolerations ensures resources are used wisely. You can configure the orchestrator to route lightweight metadata tasks to CPU instances while reserving expensive GPU nodes solely for heavy inference jobs.
- Cloud-native scaling: Because it uses native K8s Jobs, your workload scales naturally with your cluster's autoscaling policies, as well as integrates seamlessly into GitOps and IaC workflows.
(Note: The Kubernetes Orchestrator is currently in Beta for self-hosted deployments)
🛠️ Quality of life & app experience
This release includes several targeted improvements to make your daily workflows smoother.
- 100MB plugins: We’ve increased the plugin upload size limit in the App from 10MB to 100MB.
- Data quality charts: The Data Quality panel now features explicit x-axis labels on distribution charts.
- Dynamic operator inputs: For operator developers, the new ResolvableProperty allows for dynamic forms, so you can now build inputs that fetch data in real-time.
- Enterprise governance: We’ve added user_group, id, and name fields to user and viewer API queries, making it much easier for platform teams to audit permissions and integrate with identity providers.
🛡️ Security and resilience
We’ve made several improvements to the platform’s security and reliability, including:
- Clean terminations: If a pipeline operator fails or is canceled, the system recursively identifies all non-terminal child tasks and marks them as failed.
- Dependency upgrades: We’ve upgraded the following packages to resolve outstanding CVEs:
aiohttp, brotli, glob, h2, next-auth, nodemailer, starlette, and validator. - We’ve improved error handling in dataset snapshots.
- We’ve improved state transition logic for task runs.
Getting Started with 2.14.0
New to FiftyOne Enterprise? FiftyOne Enterprise is the leading visual data engine for unlocking model performance . It unifies your data, your models, and your team in one secure platform.
Get started today.
