# FLORES+-style evaluation protection: operational policy (draft)

1. Evaluation-only: prohibit use in model pretraining, continual pretraining, SFT, preference tuning, synthetic data creation or retrieval corpora used to improve systems.
2. Protect from crawlers and mirrors; require acceptance of access conditions on Hugging Face. This is an access agreement, **not a technical guarantee**.
3. Activate dataset gating from the Hugging Face Settings page; during the pilot use **manual approval** and keep records of granted access. A README `gated: true` alone is insufficient.
4. Consider holding back human post-edited references in a **separate private repository**. For strongest leakage control use hidden test sources and/or a controlled submission/scoring service; do not publish reference answers to the same broadly accessible download area.
5. Article-level permissions: arXiv and ChemRxiv do not establish one uniform license for all underlying articles. Audit each work, source excerpt and derivative translations. Seek institutional legal review before assigning the final benchmark license or enabling public access requests.
6. Version all evaluation examples and corrections, and publish checksums and aggregate score reports separately when appropriate. Do not publish raw examples in a GitHub public mirror.
7. Treat scores produced using unreviewed GPT translations as **provisional comparisons**, not reference-quality MT scores.

The policy is modeled on FLORES+ evaluation terms but is a new CRITICS policy proposal, not a copy of FLORES+'s licensing or a claim to have legally restricted third-party rights. HF gate protects only access to files for non-approved accounts, not against distribution by approved users.

References:
- https://huggingface.co/datasets/openlanguagedata/flores_plus
- https://huggingface.co/docs/hub/en/datasets-gated
- https://huggingface.co/docs/hub/en/datasets-data-files-configuration