WarBloggers-22: A Curated Dataset of War-Related Narratives from Russian Military Telegram Channels

Context & Research Significance Developed as part of a research initiative of Responsible AI for Ukraine program initiated by New York University and Ukrainian Catholic University, this dataset establishes a methodological baseline for analyzing the Russian-Ukrainian war's information space. While state-controlled media follows a rigid protocol, the unofficial "War Blogger" community (often referred to as Z-channels) provides a unique, highly emotional, and aggressive segment of the propaganda machine. This dataset captures their activity throughout the pivotal year of 2022, offering critical insights into narrative shifts, domestic mobilization, and the evolution of digital propaganda. The Data Pipeline: From Noise to Curated Content The transformation of raw social media data into a high-fidelity research artifact involved a rigorous multi-stage pipeline designed to ensure maximum thematic relevance: Initial Processing: We began with a corpus of 1,114,304 posts from 328 sources. Initial cleaning (stripping URLs, emojis, and special characters) removed approximately 240,000 entries. War-Related Filtering: To isolate content specifically related to the conflict, we developed a binary classification task. We benchmarked multiple NLP models against a manually labeled validation subset of 2,500 samples. Model Performance & Technical Edge: Our research demonstrates that specialized transformer models outperform general-purpose LLMs in this domain. Our selected model, ai-forever/ruRoberta-large, achieved superior results compared to OpenAI (gpt-5-mini) and other architectures. Classification Performance Benchmarks * 500 samples were used for test, 2000 for train Approach Precision Recall F1 ai-forever/ruRoberta-large 0.845 0.939 0.890 DeepPavlov/rubert-base-cased 0.850 0.916 0.882 OpenAI (gpt-5-mini) 0.912 0.842 0.876 TF-IDF LinearSVC 0.839 0.912 0.874 Keywords (cleaning + lemma) 0.810 0.815 0.812 Keywords (raw text) 0.841 0.604 0.703 The final curated dataset consists of 542,815 war-related posts from 221 unique bloggers. Hybrid Topic Modelling & Narrative Mapping To categorize narratives, we utilized a hybrid modeling approach that balances semantic depth with structural consistency: Frameworks: We employed BERTopic as our primary engine (using distiluse-base-multilingual-cased-v2 embeddings) and used Latent Dirichlet Allocation (LDA) as a baseline for thematic stability. Human-Refined Granularity: The topic count was fixed at k=25 to ensure stability and scientific interpretability. Each cluster was manually reviewed to ensure labels accurately reflect the aggressive linguistic style and slang of the subculture. Representative Narrative Clusters These clusters capture the multifaceted nature of the 2022 information space, ranging from frontline tactical reports to high-level geopolitical propaganda. Key examples include: International Politics & Geopolitical Relations: Focuses on the roles of the UN, EU, and global leaders (Biden, Lukashenko) in the conflict, as well as energy politics (gas) and sanctions. Keywords: UN, Europe, Washington, China, sanctions, gas, grain deal, Germany, Lukashenko. Anti-Ukrainian Narratives & Accusations of Nazism: This cluster tracks the ideological justification of the war through accusations of "Nazism" and the dehumanization of Ukrainian forces. Keywords: Nazism, crimes, nationalists, UN, Ukrainian nationalists, "Ukronazis", western media. Referendums & Annexation: Documentation of the legislative and communicative efforts surrounding the annexation of Ukrainian territories and the "partial mobilization" in Russia. Keywords: referendum, republics, partial mobilization, State Duma, decree, integration, legal right, Vladimir Putin. Dataset Artifacts & File Structure This publication includes three primary CSV files representing different stages of the research: bertopic_anonymized_war_bloggers.csv: The primary research file containing all war-related posts categorized into the 25 curated thematic clusters. lda_anonymized_war_bloggers.csv: A parallel dataset providing thematic discovery through the LDA framework for comparative structural analysis. Important: this file contains only 300k posts out of more than 540k available for BERTopic. war_related_labeled_war_bloggers.csv: The validation artifact containing 2,500 manually labeled samples used to train and evaluate the filtering models. Usage and Accessibility This corpus is intended for researchers in Political Science, Computational Linguistics, and Media Studies investigating modern information warfare. By bridging the gap between automated machine learning and qualitative discourse analysis, WarBloggers-22 provides a transparent and reproducible foundation for future studies.