> For clean Markdown of any page, append .md to the page URL.
> For a complete documentation index, see https://docs.ovalbee.com/llms.txt.
> For AI client integration (Claude Code, Cursor, etc.), connect to the MCP server at https://docs.ovalbee.com/_mcp/server.

# 0.14.0

# 0.14.0

## New

* **Dataset report checks the labels on the images.** [Dataset insights](/node-library/nodes/dataset-ops-dataset-insights) shows the AI model a few flagged frames and a few random ones with the labels and a detector's predictions side by side, so the report can tell bad labels from a weak model. The report opens with a verdict and the steps to take first, and a problem that several checks found gets one section.
* **Evaluations point at label problems.** [Evaluate object detection](/node-library/nodes/evaluate-evaluate-object-detection) and [Evaluate prelabeling efficiency](/node-library/nodes/evaluate-evaluate-prelabeling-efficiency) report confident disagreements with the labels, boxes that sit off the objects in one direction, labeled boxes no prediction comes near, and a detector that does not know the domain yet. Evaluate object detection can now feed **Findings from other nodes** in Dataset insights.
* **Grounding DINO detects the dataset's classes.** In [Predict Grounding DINO](/node-library/nodes/models-grounding-dino-predict-grounding-dino), **What to detect** has a **Dataset classes** option: the node asks for every class of the ontology the source dataset is pinned to, so a flow needs nothing but the dataset.
* **Workspace storage.** A workspace can connect storage providers and buckets of its own under **Settings > Storage**. Extra storage is a plan feature.

## Improvements

* Flow groups: drop nodes and sticky notes on a group's frame, or add them from its menu, to make them members. Frames wrap their members' edges, and edges end at the centre of a handle.
* The setup wizard of a flow template opens on its overview, can name each dataset on its own, and a template keeps a Dataset node's own name.
* Node logs show the node's own output and plain status lines, such as waiting for a machine, pulling the image, out of memory or evicted.
* Nodes with an image preview have one width, so datasets side by side are the same size.
* **Move** carries the annotations along with the images.
* Embedded public flows let the page scroll with the mouse wheel and on touch.

## Fixes

* Nodes that use the instance's AI model, such as Dataset insights, work with Claude Sonnet 5.5, Claude Opus 5.5 and Claude Fable 5.1.
* Grounding DINO predictions with classes picked from a shared ontology, such as the COCO classes, no longer fail at the end of the run.
* YOLO import reads class names written as a dictionary in `data.yaml`.
* The header of the labeling tasks list fits its width.