Researchers at CERIST in Algeria have released a fine-tuned version of the Qwen2.5-VL-3B model specifically designed for recognizing Arabic historical manuscripts. The project addresses the significant challenges posed by cursive handwriting, inconsistent ligatures, and various page artifacts like foxing and bleed-through.
The model was trained using a three-stage curriculum: pretraining on generic Arabic text, fine-tuning on manuscript data without diacritics, and final fine-tuning with diacritics restored. It employs per-manuscript-type preprocessing to handle varying pixel budgets across printed books, cursive handwriting, lithographs, and rubricated manuscripts. The current best checkpoint achieves a Character Error Rate (CER) of 6.58% and a Word Error Rate (WER) of 22.34%, outperforming benchmarks like HATFormer and the top NAKBA 2026 submission.
The team invites feedback on the demo, particularly regarding performance on difficult or unusual manuscript samples, to help improve the pipeline for Arabic Handwritten Text Recognition.