![[2602.12284] A Lightweight LLM Framework for Disaster Humanitarian Information Classification](https://arxiv.org/static/browse/0.3.4/images/arxiv-logo-fb.png){kind=logo}
[2602.12284] A Lightweight LLM Framework for Disaster Humanitarian Information Classification
Summary
This paper presents a lightweight framework for classifying humanitarian information from social media, enhancing disaster response efficiency using LLMs with minimal resources.
Why It Matters
Effective disaster response relies on timely information classification. This research addresses the challenges of deploying large language models in resource-constrained environments, offering a practical solution that can significantly improve crisis management efforts.
Key Takeaways
- Develops a lightweight framework for classifying humanitarian tweets.
- Achieves 79.62% accuracy with only ~2% of parameters using LoRA fine-tuning.
- QLoRA allows deployment with 99.4% performance at 50% memory cost.
- RAG strategies may degrade performance due to label noise.
- Establishes a reproducible pipeline for crisis intelligence systems.
Computer Science > Computation and Language arXiv:2602.12284 (cs) [Submitted on 21 Jan 2026] Title:A Lightweight LLM Framework for Disaster Humanitarian Information Classification Authors:Han Jinzhen, Kim Jisung, Yang Jong Soo, Yun Hong Sik View a PDF of the paper titled A Lightweight LLM Framework for Disaster Humanitarian Information Classification, by Han Jinzhen and 3 other authors View PDF HTML (experimental) Abstract:Timely classification of humanitarian information from social media is critical for effective disaster response. However, deploying large language models (LLMs) for this task faces challenges in resource-constrained emergency settings. This paper develops a lightweight, cost-effective framework for disaster tweet classification using parameter-efficient fine-tuning. We construct a unified experimental corpus by integrating and normalizing the HumAID dataset (76,484 tweets across 19 disaster events) into a dual-task benchmark: humanitarian information categorization and event type identification. Through systematic evaluation of prompting strategies, LoRA fine-tuning, and retrieval-augmented generation (RAG) on Llama 3.1 8B, we demonstrate that: (1) LoRA achieves 79.62% humanitarian classification accuracy (+37.79% over zero-shot) while training only ~2% of parameters; (2) QLoRA enables efficient deployment with 99.4% of LoRA performance at 50% memory cost; (3) contrary to common assumptions, RAG strategies degrade fine-tuned model performance due to labe...
