Export dataset

Dataset export

Export is scoped to one collection. OCR Yoruba uses Paddle-style line crops with train/val/test splits; VLM uses JSONL + full images.

Collection

Split ratios

100% total

Used for OCR Yoruba line-crop exports. Must sum to 100%.

Train %

Val %

Test %

Output options

Include sidecar JSON

Include metadata CSV (OCR training mode only)

OCR Yoruba ZIPs contain train/val/test line crops. VLM mode exports JSONL plus full images.