whales-identification

Compliance & licensing

EcoMarineAI touches three regulated areas: data licenses, software licenses, and environmental / ethical use. This doc consolidates everything so a legal / ethics reviewer can audit in one place.

---

1. Software license (source code)

• LICENSE — MIT, © 2024 EcoMarineAI contributors (see LICENSE in the repo root).
• Permissive: commercial, modification, distribution all allowed.
• No patent grant (MIT).

2. Training data license

The combined training corpus has two sources; the derivative model inherits the most restrictive terms of both.

2.1 Happy Whale dataset

• Source: https://happywhale.com / https://www.kaggle.com/competitions/happy-whale-and-dolphin
• Licence: CC-BY-NC-4.0 (Creative Commons Attribution-NonCommercial 4.0 International).
• Attribution required (see LICENSE_DATA.md).
• Non-commercial use only — commercial exploitation requires direct permission from Happy Whale (support@happywhale.com).

2.2 Ministry of Natural Resources RF dataset

• Source: Ministry of Natural Resources and Ecology of the Russian Federation.
• Licence: Government research-only. Not redistributable. Covered by the ФСИ grant agreement.
• Attribution format documented in LICENSE_DATA.md.

2.3 Combined effect

Use case	Permitted?	Notes
Academic research	✓	Attribute both sources
Educational use	✓	Accredited institutions
Non-profit conservation	✓	—
Scientific publications	✓	Cite both datasets
Government monitoring (RF)	✓	With ФСИ approval
Commercial products	✗	Blocked by CC-BY-NC and gov restrictions
Open-source tools (MIT)	✓	For non-commercial downstream use
Startups / for-profit orgs	✗	Unless they obtain commercial licence

3. Pre-trained model licenses

Component	Upstream	Licence
OpenCLIP ViT-B/32	laion/CLIP-ViT-B-32-laion2B-s34B-b79K	Apache 2.0
EfficientNet-B4 (ImageNet)	timm efficientnet_b4 pre-trained weights	Apache 2.0
ArcFace head (fine-tuned)	ktakita/happywhale-exp004-effb4-trainall	MIT (Kaggle user content — CC0/MIT mixture)
ResNet-101 (legacy)	baltsat/Whales-Identification	MIT
rembg background removal	danielgatis/rembg	MIT

All permissive, but each component’s attribution requirement is preserved in LICENSE_MODELS.md.

4. Hugging Face mirror

0x0000dead/ecomarineai-cetacean-effb4 on HF carries the combined licence: CC-BY-NC-4.0 (taking the strictest of the inputs). The HF model card lists all upstream sources in the datasets front-matter and in the ## Licensing section.

5. Third-party dependency audit

See LICENSES_ANALYSIS.md for the full 159-dependency license breakdown. Summary:

• 99.4 % permissive (MIT / BSD / Apache 2.0 / MPL).
• 0.6 % LGPL (e.g. libmysqlclient transitively through aiomysql) — acceptable.
• 0 % GPL / AGPL — deliberately excluded.

6. GDPR / personal data

• The service does not log input images by default.
• The drift monitor stores only aggregate statistics (rolling window of floats), no images.
• The SQLite/Postgres sinks store filename + predictions — no raw pixel data.
• If you bolt on image storage (for active-learning workflows), make sure to inform users and get consent for their photos being retained.
• No PII in training data (the individual_id column refers to whales, not people).

7. Environmental impact

• Training: one-off ~5 GPU-hours for EffB4 fold 0 → ~5 kg CO2 equivalent at typical data-centre mix. Documented in DOCS/RESEARCH_NOTES.md.
• Inference: ~490 ms CPU per image. A typical 1 000-photo survey uses ~8 minutes of CPU wall time → negligible.
• We do not recommend running the service 24/7 on an always-on GPU. Use CPU autoscaling.

8. Dual-use / ethical considerations

• The system is designed for conservation and research, not commercial exploitation of marine resources.
• It could theoretically be repurposed for whaling operations targeting specific individuals. We mitigate this by:
  • Documenting prohibited uses in LICENSE_MODELS.md (“Applications that harm marine mammals”, “Surveillance for hunting purposes”).
  • Binding the model to the non-commercial CC-BY-NC-4.0 upstream.
  • Refusing to provide a “species → habitat → best hunting coordinates” integration.
• Researchers who suspect misuse should contact the team and, separately, the upstream data providers.

9. ГОСТ 7.32-2017 alignment

The source code alignment is complete. The accompanying research report (НТО) must still follow ГОСТ rules for its Russian-language deliverable. A non-exhaustive checklist of what the code side enables:

ГОСТ requirement	Source artifact
Библиографические ссылки на методы	RESEARCH_NOTES.md §6
Воспроизводимость результатов	scripts/compute_metrics.py, scripts/benchmark_*
Структурированное описание архитектуры	ML_ARCHITECTURE.md
Документация пользовательского интерфейса	USER_GUIDE_BIOLOGIST.md
Документация API	API_REFERENCE.md
План тестирования	TESTING_STRATEGY.md
Развёртывание и эксплуатация	DEPLOYMENT.md, MLOPS_PLAYBOOK.md

10. Contact for compliance questions

For data-license questions: the upstream providers (Happy Whale and the Ministry of Natural Resources RF). For code-license questions: open a GitHub issue tagged legal.

This site is open source. Improve this page.