TRACE-BN: Transferring Bangla-English Tutoring Behavior to a Sub-1B Offline Language Model

This paper introduces TRACE-BN, a curriculum-guided bilingual tutoring dataset of structured traces for Bangla-to-English language learners at the CEFR A1–A2 level. Each trace encodes a complete seven-field pedagogical sequence combining word-level glosses, literal and natural translations, contrastive Bangla grammar explanations, common learner mistakes, and targeted practice with answers. We show that a compact sub-1B offline model (Qwen3-0.6B) can acquire this multi-component tutoring behavior via 4-bit LoRA fine-tuning, improving schema validity from 85.4% to 95.8% while raising translation and pedagogical scores. We validate the approach through an expert human audit of the supervision signal and a reference-aware dual-judge evaluation showing consistent improvements across all evaluated tutoring dimensions.