TabuLM: Morphology-Aware Pre-training for Kinyarwanda Tabular Data

TabuLM integrates row, column, and cell-type embeddings with learned table-structure attention to improve tabular representation learning in morphologically rich languages. The model uses Masked Cell Recovery and Column Type Prediction objectives, pre-trained on 172 Kinyarwanda government tables.