Hello CC-OCR-V2 maintainers,
We are evaluating the recognition subset of CC-OCR-V2 for use as an OCR benchmark authority corpus. Before downloading or using the asset bytes, we need to resolve the distinction between the repository's MIT software license and the rights attached to the dataset images and annotations.
At the pinned dataset revision f80226d958d748daa0bf805ed6f69c2d41823e80, the Hugging Face card declares license: mit, while the GitHub repository contains an MIT license referring to software and associated documentation. We could not find a standalone data-asset license or a per-asset provenance manifest.
Could you please clarify or publish the following?
1. Dataset asset license
Please state whether the license covers the dataset images, questions, answers, and annotations, not only the repository software/documentation.
It would help to explicitly state whether the following uses are permitted:
- download and local storage;
- format conversion and deterministic preprocessing;
- OCR evaluation and benchmark comparison;
- publication of aggregate metrics and reproducibility metadata;
- commercial/internal evaluation;
- redistribution of original assets, derived assets, or neither.
A dedicated DATA_LICENSE.md would make this boundary unambiguous.
2. Per-asset provenance
For each distributed image/annotation pair, please provide a machine-readable manifest such as ASSET_PROVENANCE.csv or JSONL containing, where applicable:
- CC-OCR-V2 relative asset path or stable asset id;
- original dataset/source name;
- original asset id and source URL;
- copyright holder or provenance statement;
- original license identifier and license URL;
- redistribution status;
- annotation source;
- upstream source-document id or grouping key.
3. Source-document grouping
For leakage-safe benchmark role assignment, image hashes alone are insufficient. Multiple images may originate from the same document, template, scene, or upstream collection unit.
Please provide a stable source-document/group identifier that can be used to keep calibration, candidate, and sealed-holdout roles independent.
4. Recognition subset
Our current candidate scope is limited to:
- recognition/multi_lingual_recognition;
- recognition/natural_scene_recognition.
We are not treating extraction, grounding, parsing, or QA answers as canonical OCR transcription truth. If licensing/provenance differs by task subtree or upstream source, please document that distinction.
Until these facts are available, we will keep acquisition and benchmark-authority use blocked rather than infer per-asset rights from the aggregate MIT metadata label.
Thank you for clarifying the intended data rights and provenance boundary.
Hello CC-OCR-V2 maintainers,
We are evaluating the recognition subset of CC-OCR-V2 for use as an OCR benchmark authority corpus. Before downloading or using the asset bytes, we need to resolve the distinction between the repository's MIT software license and the rights attached to the dataset images and annotations.
At the pinned dataset revision f80226d958d748daa0bf805ed6f69c2d41823e80, the Hugging Face card declares license: mit, while the GitHub repository contains an MIT license referring to software and associated documentation. We could not find a standalone data-asset license or a per-asset provenance manifest.
Could you please clarify or publish the following?
1. Dataset asset license
Please state whether the license covers the dataset images, questions, answers, and annotations, not only the repository software/documentation.
It would help to explicitly state whether the following uses are permitted:
A dedicated DATA_LICENSE.md would make this boundary unambiguous.
2. Per-asset provenance
For each distributed image/annotation pair, please provide a machine-readable manifest such as ASSET_PROVENANCE.csv or JSONL containing, where applicable:
3. Source-document grouping
For leakage-safe benchmark role assignment, image hashes alone are insufficient. Multiple images may originate from the same document, template, scene, or upstream collection unit.
Please provide a stable source-document/group identifier that can be used to keep calibration, candidate, and sealed-holdout roles independent.
4. Recognition subset
Our current candidate scope is limited to:
We are not treating extraction, grounding, parsing, or QA answers as canonical OCR transcription truth. If licensing/provenance differs by task subtree or upstream source, please document that distinction.
Until these facts are available, we will keep acquisition and benchmark-authority use blocked rather than infer per-asset rights from the aggregate MIT metadata label.
Thank you for clarifying the intended data rights and provenance boundary.