BhashaLLM: A QLoRA-Based Framework for Bangla Text Generation and Handwritten Character Recognition

Abstract

Bangla is spoken by more than 250 million people but remains poorly served by mainstream large language models (LLMs), which are typically trained, tuned, and benchmarked with English or high-resource languages in mind. This paper presents BhashaLLM, a framework that pairs a QLoRA-fine-tuned instruction model with a vision-language OCR pipeline for handwritten Bangla, built entirely around a single 16 GB consumer GPU. We benchmark nine LLM and vision-language architectures, from 1.5B to 12B parameters, spanning the Qwen, Gemma, Mistral, and Llama families, on translation, summarisation, and OCR-correction tasks. Two failure modes dominate the smaller models: language reversion, where the model answers in English despite a Bangla prompt, and script confusion, where Bangla text is interleaved with Hindi Devanagari characters, affecting up to 23% of Qwen-1.5B’s outputs. Llama-3.2-11B is identified as the strongest candidate, holding correct Bangla script in over 99% of generations and scoring highest (14/15) on a blind human evaluation of translation quality. For handwritten OCR, a Qwen-VL-based model fine-tuned with QLoRA on a self-collected 1,500-image dataset lowers the character error rate from 28% to 12%, placing it close to Google Cloud Vision’s reported 13.89% CER on a separate handwritten Bangla benchmark. We describe the memory-efficiency techniques, 4-bit NF4 quantisation, low-rank adaptation, and gradient checkpointing, that make an 11-billion-parameter model trainable on 16 GB of VRAM, and report the failure modes and open questions, including script confusion in smaller models and an unresolved inter-rater reliability measurement, that remain after fine-tuning.