For AI agents: a documentation index is available at the root level at /llms.txt. Append /llms.txt to any URL for a page-level index, or .md for the markdown version of any page.
Public templates.Save as template can publish a template to every workspace of the installation with Public template, and can replace one of your existing templates instead of adding a new one. Tags, a Markdown readme and publishing sit under Advanced settings. Official templates now come live from the official OvalBee library, so a new one appears in every installation without an update. See Templates and sharing.
Machine classes. Administrators name the machines node runs and deployments can use under Machine classes in the platform settings, and a node or a deployment picks one of them. See Compute.
Import from cloud storage. The Import from cloud storage node imports images or videos with their annotations from an S3-compatible bucket and detects the annotation format.
Label train, validation and test. The Label train, validation and test node gives each labeler one queue that works through the three sets in order, and each set’s submissions land in a labeled dataset of its own.
Active learning with live training.Prelabel with live training pre-labels a dataset with the current weights of a running live training session without stopping it, and can accept images whose boxes are all confident enough. Active learning curve plots mAP against the number of labeled images.
Group styles. Flow settings offer three styles for groups, and templates keep a flow’s groups.
Running deployments. The deployments tab filters by All or Running.
Improvements
Official library URL in the platform settings names the OvalBee instance that official templates and live examples come from. A value that still points at the old catalog manifest keeps working.
Labeling and Review nodes refresh their queues as soon as work arrives. Their scheduled run, now every 15 minutes instead of every minute or two, returns expired claims.
Train/validation split keeps every image of its previous split on rerun and places only new ones, and has an Add a test split switch. Random sample names its dataset after its source.
Sample images similar to failures has an optional Exclude images in dataset, so images already labeled are not sampled again.
Grid thumbnails draw boxes in the classes’ ontology colors, and a class without a color gets the same one every time.
Evaluation reports leave out findings charts and tables that have nothing to show.
Collaborators no longer see Storage, Models and Ontologies. Sync from disk in the node picker is shown only to instance administrators.
Fixes
Embedding map thumbnails no longer go blank a day after their run.
Review, Consensus review and other steps that cannot edit labels no longer offer Smart Tool or Live Training in the labeling tool, and a rejected image is no longer sent to live training.
A Condition node compares numbers as numbers, where a numeric comparison used to stop the flow.
A dataset card no longer says “No assets” after a node refreshed its statistics.
The CLIP index badge no longer drops to 0 images when a scheduled rerun finds the dataset unchanged.
The labeling tool no longer fails to load an image whose pre-label carries a confidence tag.
A template saved without tags in an update keeps the tags it had.
