New dataset enables small offline models for Bangla-English tutoring

Researchers have developed TRACE-BN, a dataset and method for transferring Bangla-English tutoring capabilities to a small, offline language model. The system uses curriculum-guided structured tutoring traces, generated by Gemini 3.5 Flash Lite, to teach English to Bangla speakers at a beginner level. This behavior is then transferred to a sub-1B model, Qwen3-0.6B, using LoRA with 4-bit quantization, significantly improving its performance on translation, grammar explanation, error diagnosis, and practice alignment. AI

IMPACT Enables resource-constrained environments to deploy specialized AI tutors for language learning.

RANK_REASON This is a research paper detailing a new dataset and method for transferring AI capabilities. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.CL →

• Bangla
• CEFR A1-A2
• English
• Gemini 3.5 Flash Lite
• LoRA
• Qwen3-0.6B
• TRACE-BN

• paper
• model release
• product

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →