Datasets:
Convert to parquet (loading scripts are no longer supported)
#2
by davanstrien HF Staff - opened
datasets no longer supports dataset loading scripts, so load_dataset('biglam/yalta_ai_tabular_dataset') fails with RuntimeError: Dataset scripts are no longer supported.
This PR converts both configs (YOLO, COCO) and all three splits to parquet, and removes the loading script. Boxes are derived from the source YOLO labels exactly as the script did: YOLO keeps [x_center, y_center, w, h] in absolute pixels, COCO uses [min_x, min_y, w, h] with area and iscrowd. Images are embedded, so loading no longer depends on Zenodo.
Two defects in the original COCO config are fixed here:
- annotation
idwas passed the image id, so every box on a page shared one id. Ids are now unique per annotation. image_idwas declaredstringbut populated with an int; it is nowint64to match the value.
Source data unchanged: https://doi.org/10.5281/zenodo.6827706
davanstrien changed pull request status to merged