Datasets:
Tasks:
Image-Text-to-Text
Formats:
parquet
Languages:
Portuguese
Size:
< 1K
ArXiv:
Tags:
ocr
benchmark
document-understanding
brazilian-portuguese
text-recognition
handwriting-recognition
License:
| # PRIVACY_NOTICE | |
| ## Purpose | |
| This benchmark package may include or reference document images, cropped images, transcriptions, and metadata used solely for OCR evaluation and benchmarking. | |
| ## Privacy rule for inclusion | |
| Documents containing personal data may be included in the benchmark **only if** at least one of the following is true: | |
| 1. the document has been **effectively anonymized** so that the relevant person is no longer identified or identifiable in the benchmark asset made available; or | |
| 2. there is a **clear and documented legal basis** for public disclosure and redistribution of the relevant benchmark asset. | |
| If neither condition is met, the document must not be redistributed as part of the public benchmark package. | |
| ## Minimum review before publication | |
| Before a document is included in the benchmark package, Dharma-AI should confirm and record: | |
| - source URL and source organization; | |
| - document type and context of publication; | |
| - whether the source document is publicly accessible; | |
| - whether the benchmark asset reproduces personal data; | |
| - whether the personal data include any higher-risk categories; | |
| - whether anonymization has been performed and validated; and | |
| - the final release decision in `SOURCES_LEGAL.md`. | |
| ## Examples of personal data to review | |
| The review should look for, at minimum: | |
| - names linked to identifiable persons; | |
| - CPF, RG, OAB, passport, or similar identifiers; | |
| - signatures; | |
| - addresses, e-mails, phone numbers; | |
| - photographs or biometrics; | |
| - financial or employment information; | |
| - medical or other specially protected information; and | |
| - combinations of fields that make re-identification reasonably possible. | |
| ## Anonymization standard | |
| Anonymization should be applied to the **benchmark asset actually released**, not only to internal copies. Depending on the case, this may require: | |
| - redaction or masking in the image/PDF; | |
| - removal of identifying metadata; | |
| - exclusion of specific pages or regions; or | |
| - release of transcription-only or metadata-only variants. | |
| If effective anonymization cannot be achieved without undermining legal or privacy requirements, the item should not be publicly redistributed. | |
| ## Public official documents | |
| Even where a source document is publicly accessible or may qualify as an official act for copyright purposes, that does **not automatically eliminate privacy review**. Public availability alone is not a blanket authorization to redistribute benchmark-ready copies in a machine-readable package. | |
| ## Redistribution fallback | |
| Where there is uncertainty about privacy or redistribution rights, the default fallback is: | |
| - publish the benchmark framework, metrics, and metadata; | |
| - retain only links, hashes, or identifiers for the underlying source; and | |
| - withhold the original image/PDF/transcription from the public package. | |
| ## Contact and review | |
| Questions about whether a source may be included should be escalated to legal review before publication. | |