DATA_AND_THIRD_PARTY_NOTICE.md

# Data and Third-Party Notice

## Scope

The `urnova-95m` model was trained from scratch. It was not initialized from
mBERT, XLM-R, DunbaaBERT, HPLT-BERT Urdu, or another third-party model
checkpoint.

The original urnova model weights and locally generated tokenizer artifacts
are distributed under the Apache License, Version 2.0, to the extent of the
rights held by the urnova publisher.

This licence does not relicense the underlying pretraining documents.

## Training-data source

The production corpus was derived from the Urdu Arabic-script portion of
HPLT 3.0.

HPLT states that:

- it does not own the text from which the dataset was extracted;
- it licenses the packaging of the dataset under Creative Commons CC0;
- users remain responsible for compliance with applicable law.

Official dataset identifier:

`HPLT/HPLT3.0`

No raw or reconstructed HPLT documents are included in this model repository.

## Recorded production provenance

The frozen production lineage records:

- source documents: 1,921,027;
- source-token count: 1,246,973,570;
- production training tokens processed: approximately 4.0 billion;
- source corpus name in the deduplication manifest: `hplt3`;
- source file SHA-256:
  `b4ede7d4548fc3db385666eddd796ae2ceb0d37788bbe85a13ff8abd81ebb991`;
- tokenized output SHA-256:
  `b07ad6d48efe3e5147d875b09acf4a53ce675d078e3d868e8d778e15c8c863aa`;
- upstream corpus identity SHA-256:
  `0ba4bf10ad09c3b8105f8ac20dd62953e8537664b8cda0ea4557a6ec275817d3`;
- tokenizer vocabulary SHA-256:
  `3acb94ea6bca8b2a0256d9d46c545bb9c033977c71671f8bc52cb1aa6ada92bc`.

Phases A, B, and C used the same tokenized production corpus with different
packing and context-length policies.

## Provenance limitations

The local production records do not preserve an immutable HPLT upstream
revision or acquisition timestamp.

The tokenizer-training manifest records 199,117 documents from the local file
`dedup-200k.jsonl`, with SHA-256:

`75d73d7d874525ab9c076f62ebbeb6bfd7bd0b63acac68bee2db657771855f34`

That tokenizer manifest does not independently record the original dataset
identifier or immutable upstream revision.

These are known provenance limitations and must not be represented as fully
resolved source-level provenance.

## Third-party rights

Apache-2.0 applies only to rights held by the licensors in the distributed
urnova artifacts. It does not grant permission to use:

- copyrighted text that may be reproduced from training data;
- names, logos, or trademarks appearing in source material;
- personal information or content subject to privacy or publicity rights;
- material restricted by website, publisher, contractual, or database terms.

Users are responsible for determining whether their intended use complies
with applicable law.

## Takedown and concerns

For concerns relating to this model repository, open a discussion at:

https://huggingface.co/mahwizzzz/urnova-95m/discussions

Provide enough information to identify the affected material and explain the
basis of the request.

Valid requests will be investigated. Where appropriate, affected sources will
be excluded from future training runs, and the model card will identify which
previous model versions remain affected.

HPLT publishes its own dataset notice-and-takedown procedure separately.