CREATING A CRISIS-DEPENDENT DATASET FOR ADAPTIVE IRM
DOI:
https://doi.org/10.34185/1991-7848.itmm.2026.01.073Keywords:
context-aware generation, large language models, multimodal dataset, crisis data annotation, crisis informatics, HumAID dataset, context injection, model behavior adaptationAbstract
In crisis communications, Large Language Models (LLMs) have the potential to assist in generating guidance and recommendations; however, their default behavior often ignores the specific nature of the event. This reduces relevance and may pose risks in critical situations. This paper presents an approach to constructing a specialized dataset for training and evaluating Adaptive IRM - a module that injects latent crisis context into the forward pass of an LLM. The HumAID corpus of disaster-related tweets was used as a foundation, with abstract questions generated without explicit mentions of the crisis type. The resulting dataset (~41K examples) enables the assessment of whether models equipped with Adaptive IRM can produce responses that vary according to the crisis type, thereby improving both relevance and safety.
References
Otal H. T., Canbaz M. A. C. LLM-Assisted Crisis Management: Building Advanced LLM Platforms for Effective Emergency Response and Public Collaboration // 2024 IEEE Conference on Artificial Intelligence (CAI). 2024. P. 851–859. DOI: 10.48550/arXiv.2402.10908.
Fengyi X. та ін. Large language model applications in disaster management: An interdisciplinary review // International Journal of Disaster Risk Reduction. 2025. Vol. 127. Art. No. 105642. DOI: 10.1016/j.ijdrr.2025.105642.
Bereziuk M. O., Huda A. I. Kontekstno zalezhna adaptatsiia vidpovidei heneratyvnykh LLM. Informatsiini tekhnolohii v metalurhii ta mashynobuduvanni – ITMM2025 : tezy dop. Mizhnarodnoi nauk.-tekhn. konf. (m. Dnipro, 23-24 bereznia 2025 r.). Dnipro, 2025. C. 503–508. DOI: 10.34185/1991-7848.itmm.2025.01.089.
Alam F., Qazi U., Imran M., Ofli F. HumAID: Human-Annotated Disaster Incidents Data from Twitter with Deep Learning Benchmarks [Elektronnyi resurs]. 2021. URL: https://arxiv.org/abs/2104.03090 (data zvernennia: 11.03.2026).
