CrimeaIsUkraineOrg/crimea-sovereignty-corpora

Crimea Digital Sovereignty: Training Corpora Sovereignty Framing

Analysis of how LLM training corpora frame Crimea's sovereignty. A Rust-based classifier with 90 legal-grounded signals scanned 34.1M Crimea-mentioning documents across three C4 language splits.

Description

Documents from Google's C4 corpus that mention Crimea, classified by sovereignty framing using a deterministic regex classifier (no ML/learned parameters). 90 signals across English, Russian, and Ukrainian, each grounded in legal provenance (OFAC SDN, EU Regulations, GEC reports).

Files

• data.parquet — C4-English Crimea mentions (0 documents)
• sovereignty_training.parquet — Curated sovereignty training examples (0 examples)

Key Findings (full C4 census)

Split	Total docs	Russia-framing	%
English	286,117	3,607	1.26%
Ukrainian	3,639,461	6,271	0.17%
Russian	30,207,220	881,644	2.92%
Total	34,132,798	891,522	2.61%

• 95.3% of Russia-framing documents originate from independent (non-state) sources
• Only 4.7% from state-controlled or sanctioned sources (OFAC/EU/UK lists)
• Validated by two independent annotators on 300 samples (100 per split): weighted precision 93.9%

Source classification

Source tier	Count	%	Legal provenance
Independent	849,761	95.3%	No state ties identified
State-adjacent	16,772	1.9%	Sberbank/Gazprom, EU Pkg 16
State media (T1)	14,406	1.6%	OFAC SDN, EU Regulations
Sanctioned proxy	4,928	0.6%	GEC 2020, OFAC EO14024
Government	3,156	0.4%	Russian federal law
State-controlled (T2)	2,470	0.3%	Gazprom Media/NMG
Pravda network	29	<0.1%	VIGINUM/SGDSN 2024

Classifier source code

github.com/IvanDobrovolsky/crimeaisukraine/c4_sovereignty/scanner

Citation

@misc{dobrovolskyi2026digital,
  author = {Dobrovolskyi, Ivan},
  title = {Digital Annexation: A Computational Audit of Crimea's Sovereignty Framing in Large Language Models},
  year = {2026},
  url = {https://crimeaisukraine.org}
}

License

CC-BY-4.0