Saad A.

Artificial Intelligence Specialization · Cohort 2026

DiacriticS: Fine-Tuning a Language Model for Arabic Diacritisation on SadeedDiac-25 Benchmark

DiacriticS is a research project focused on supervised fine-tuning (SFT) of open-weights language models to achieve robust Arabic Text Diacritisation (ATD) under strict contamination control. The project utilises a training corpus derived from Sadeed Tashkeela (~53 million words) with minimal overlap (0.4%) against evaluation data, while benchmark evaluations are conducted on SadeedDiac-25, a contamination-controlled dataset of 1,200 curated, human-reviewed paragraphs evenly split between Modern Standard Arabic and Classical Arabic. Baseline zero-shot evaluations measured using formal Diacritic Error Rate (DER) and Word Error Rate (WER) metrics reveal that sentence-final case endings (I'rab) and Classical Arabic domains remain the primary sources of error, establishing the project's next steps to execute SFT and perform targeted error analysis to resolve syntactic context bottlenecks.

Read the final report (PDF)