ahammadshawki8/Shawki_Codes_IUTDL ? reverse-engineered prompt

Build me a notebook and scripts for the Bengali Hallucination Detection challenge.

I want a system that takes a Bengali prompt and a candidate answer, then predicts whether the answer is faithful or hallucinated. Please make it work in a layered way, first checking for exact or fuzzy matches against known public source data, then using Bengali aware text matching for idioms and similar cases, and only using a local LLM judge for the hard leftovers. It should be able to run end to end for a Kaggle style submission, with a clear path from raw data to final CSV output.

If possible, include the pieces for source matching, fuzzy multiple choice matching, simple retrieval for context based QA, and a small local evaluation flow so I can test accuracy before submitting. Use the data and cached files already in the repo, and look up current docs online if you need to.