Large, labeled multilingual OCR corpora are rare and expensive to produce. This dataset addresses that gap by providing over 12.25 million synthetically generated samples across six language variants, formatted as HDF5 shards with structured annotations (word/line/paragraph bboxes, quads, and reading-order graphs). The scale and structured labels make it practical to train and evaluate both detection and recognition components together, and it was used to train NVIDIA's Nemotron OCR v2.
What Sets It Apart
- Scale and multilingual coverage — 12,258,146 samples across English, Japanese, Korean, Russian, Simplified and Traditional Chinese, stored as per-language train/val/test HDF5 shards. So what: enables robust multilingual training and joint detection+recognition experiments without repeatedly stitching smaller datasets.
- Structured, multi-level annotations — each sample includes word_bboxes, line_bboxes (with para_idx/line_idx/word_indices), para_bboxes, and a relation_graph encoding reading order. So what: supports fine-grained supervision for line/word-level recognition, reading-order modeling, and relation-aware post-processing.
- HDF5 shard layout and metadata — JPEG-encoded images, per-sample annotations as JSON strings, dimensions, labels, qualities and sample_ids are included. So what: optimized for high-throughput I/O in training pipelines and straightforward to load with standard Python/HDF5 stacks.
- Built for model training — explicitly created to train Nemotron OCR v2 and integrated into NVIDIA NeMo tooling. So what: provides a reproducible training source for models that require joint detection, recognition, and layout reasoning.
Who it's for and trade-offs
Great fit if you are a researcher or engineer training multilingual OCR models, evaluating reading-order or layout-aware recognition, or benchmarking detection+recognition stacks at scale. Look elsewhere if you need in-domain, photographed real-world scans for a narrow vertical — synthetic renderings may not capture some real-world noise, camera artifacts, or rare fonts. Also plan for storage and I/O: the full corpus occupies multiple terabytes (reported ~5.45 TB) and requires tooling to stream HDF5 shards efficiently. Licensing (CC BY 4.0) permits commercial use but requires attribution.
Where it fits
Use this dataset as a large pretraining or augmentation source to reduce annotation costs and accelerate model convergence; combine with smaller, domain-specific real datasets for final fine-tuning to address domain-shift.
