Automated MCQ Generation with Options and Answering based on Text and Images using NLP

My final year project: a system that auto-generates Bangla-language multiple-choice questions from text and images. I fine-tuned and compared multiple LLMs (Qwen3-VL-8B, BanglaT5, Gemma, Pixtral/Ministral 12B) using LoRA/PEFT on the Noirbittik Bangla dataset via Kaggle — BanglaT5 proved best for explainability, Qwen3-VL-8B for generation quality. I built a Streamlit UI with quiz interaction, Jaccard-similarity checks against training data to catch leakage, history/records tabs, screen recording, and CSV/JSON export, deployed via Cloudflare tunnel. I also fixed several critical training bugs (data leakage, global variable misuse, missing eval metrics, checkpoint selection) and separately debugged a Mistral-7B notebook through SFTTrainer/FastLanguageModel compatibility issues.

Note: the repo mcqcreator is currently empty on GitHub — you'll want to push your code before linking it, or use the repo where the actual work lives if it's elsewhere.