Applied machine learning / 2025
Multimodal disaster classification
Classifying disaster-related posts on CrisisMMD using text, images and fine-tuned vision-language models.
Task and dataset
The project classifies disaster-related social posts using the CrisisMMD dataset. Experiments include text-only, image-only and combined text–image inputs. It covers whether a post contains useful information and the type of humanitarian information it contains. The work was published at ASONAM 2025 with me as first author.
Models and experiments
I led the project, working with Gemma and Llama models, prompt design and parameter-efficient fine-tuning. With Gemma 3-4B, label refinement and prompting raised reported informativeness accuracy from approximately 73% to 85.53%.
Fine-tuning Llama 3.2-11B Vision with LoRA used both text and images. The reported accuracies were 94.78% for informativeness and 91.62% for humanitarian classification. These figures refer to those specific project tasks on CrisisMMD.
Experiment tooling
Python scripts generated experiments with different hyperparameters and coordinated training jobs across multiple GPUs. PyTorch, Transformers and scikit-learn supported model training and analysis.