Fine-Tuned Lightweight Model Grades Bangla Written Answers Semantically

A bilingual Bangla-English grading system fine-tunes a lightweight LLM on a synthetic dataset to score student answers using question, reference, and student response as input, producing numeric scores and feedback. The approach prioritizes semantic correctness over lexical overlap for low-resource educational deployment.