license: other
task_categories:
- image-to-text
- document-question-answering
language:
- zh
pretty_name: Taiwan Traditional Chinese OCR Filtered PaddleOCR-VL Labels
tw-ocr
Filtered OCR/document parsing records for Traditional Chinese OCR experiments.
The public Hugging Face upload is parquet-only for dataset rows. Every parquet
row embeds non-empty image bytes and is written with Hugging Face datasets
Image feature metadata. The physical parquet storage is the standard Image
struct ({bytes, path}) with image.path = null; when loaded through
datasets, the image column is an Image feature instead of a generic dict
column. The source path is retained only in the separate image_path
provenance column.
Human, synthetic, dataset, and teacher labels remain separated in each
record_json payload. Public exports use the accepted split only;
rejected/manual-review rows are local audit artifacts and must not be uploaded
to the public dataset.
| split | records | rows with embedded image.bytes | rows missing image.bytes | embedded image bytes |
|---|---|---|---|---|
| accepted | 75903 | 75903 | 0 | 3480370760 |
Post-write image-byte verification:
- total rows: 75903
- rows with non-empty
image.bytes: 75903 - rows missing
image.bytes: 0 - rows with empty
image.bytes: 0 - rows with path-only images: 0
- rows with public
image.path: 0 - rows with image byte length mismatch: 0
- parquet files with HF
Imagemetadata: 8 - parquet files missing HF
Imagemetadata: 0 - valid: True
License
Licenses are per row in the license field. This mixed dataset contains
synthetic records and third-party datasets with non-commercial or research
constraints. Users must filter rows according to their allowed use.