tw-ocr / README.md
voidful's picture
Upload accepted tw-ocr parquet shards with embedded image bytes
ec7bfee verified
|
Raw
History Blame Contribute Delete
1.31 kB
---
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 public
row must embed non-empty image bytes as a Hugging Face `Image` struct
(`{bytes, path}`). The local gzip JSONL files are export intermediates for
provenance and are not canonical public data files; do not upload them as HF
dataset rows.
Human, synthetic, dataset, and teacher labels remain separated in each
`record_json` payload. Public Hugging Face exports include accepted rows only;
rejected/manual-review rows are local audit artifacts and should not be pushed
to the public dataset.
Source refilter manifest: `experiments/manifests/refilter_label_agreement_v42_submit1_latest.json`
| split | records | rows with source image path | rows missing source image path |
| --- | ---: | ---: | ---: |
| accepted | 54196 | 54196 | 0 |
## 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.