Authors - The Quan Trong, Nguyen Trong Nhan Abstract - The integration of large language models (LLMs) into primary educa tion remains limited in low resource, diglossic languages like Sinhala. General purpose models often produce grammatically inconsistent or cognitively over whelming output for young learners. This paper introduces a grade-adaptive, con straint-driven framework for automated Sinhala story and quiz generation target ing Grades 1-5. Building upon an 8-billion-parameter Sinhala-adapted LLaMA 3 model, we apply Quantized Low-Rank Adaptation (QLoRA) using a curated multi-task educational dataset. The system enforces tier-specific linguistic con straints separating conversational Sinhala for lower grades from formal written Sinhala for upper grades while embedding strict structural rules such as con trolled sentence counts (5-6 vs. 7-8) and validated multiple-choice formats (3 vs. 4 options). Evaluation on 100 structured prompts demonstrated substantial im provements over a zero-shot baseline: structural compliance increased from 64% to 93%, and hallucination-related failures decreased from 31% to 8%. Further more, evaluation against 50 unseen real-world classroom prompts yielded a 0.0% crash rate and 95% register adherence, confirming robust qualitative perfor mance. Results demonstrate that diglossia-aware dataset engineering and con straint-aware fine-tuning enable reliable, pedagogically aligned deployment of LLMs in low-resource primary learning environments.
Open Zoom