Reference for ultralytics/models/yolo/world/val.py#

class WorldValidator(DetectionValidator):
    """A validator for YOLO-World models that sets dataset class names before validation.

    Open-vocabulary YOLO-World models default to 80 COCO classes, so validating on a dataset with different classes
    (e.g. LVIS) fails or yields zero metrics. This validator generates text embeddings for the dataset's class names so
    standalone `model.val()` works. During training, classes are set via the `on_pretrain_routine_end` callback instead.
    """

def __call__(self, trainer=None, model=None):
    """Set dataset classes for standalone validation, then run validation."""
    if trainer is None:  # standalone val; training sets classes via on_pretrain_routine_end callback
        self.device = select_device(self.args.device, verbose=False)
        if not isinstance(model, torch.nn.Module):
            from ultralytics.nn.tasks import load_checkpoint

            model = load_checkpoint(model or self.args.model, device=self.device)[0]
        model.eval().to(self.device)
        self.args.data = convert_ndjson_to_yolo_if_needed(self.args.data)  # match BaseValidator dataset handling
        names = [name.split("/", 1)[0] for name in check_det_dataset(self.args.data)["names"].values()]
        current = model.names.values() if isinstance(model.names, dict) else model.names  # names may be a list
        if list(current) != names:  # regenerate prompts only if class order differs from dataset
            state = (model.names, model.txt_feats, model.model[-1].nc)  # restore after to avoid leak to caller
            model.set_classes(names, cache_clip_model=False)
            model.names = dict(enumerate(names))  # set_classes updates embeddings/nc but not names
            try:
                return super().__call__(trainer, model)
            finally:
                model.names, model.txt_feats, model.model[-1].nc = state
    return super().__call__(trainer, model)