Ultralytics Platform & YOLO26: train and export a model

This tutorial follows a browser-based object detection workflow in Ultralytics Platform, from cloning an annotated construction PPE dataset to testing a deployed API. The presenter uses YOLO26 Nano and configures a 100-epoch training run on an RTX Pro 6000 GPU. The training interface includes batch size, image size and advanced parameters, plus an estimated training cost.

The dataset has 11 classes, with labels for people and protective equipment such as helmets, vests and gloves. The presenter reviews class distribution and dataset splits before training, notes an imbalance, and recommends more data with balanced classes for an actual project. After training, he checks precision, recall, mAP and loss charts, then tests predictions on sample images. Those examples show detections, but do not establish accuracy across other data.

For on-device use, the relevant section is model export. The presenter selects OpenVINO, starts the conversion and downloads the exported model. He discusses other export formats, though the tutorial does not demonstrate inference on local hardware.

The final section deploys the model to Ultralytics cloud and tests image predictions through its endpoint. The interface provides API-key-based code examples for Python and JavaScript. The demonstrated training and deployment use hosted services; the downloaded export is the part viewers can take into a separate local AI workflow.