Fine-tuning MuRIL for Multilingual Citizen Grievance Classification (with all the bugs I hit)

Short summary

A detailed writeup of fine-tuning Google's MuRIL model to classify Indian citizen grievances across 14 government departments in Hindi, Hinglish, and English. The author honestly documents four bugs encountered—including a misleading 100% accuracy from train/test leakage, RNG state contamination across categories, and keyword bleed—achieving 87.3% accuracy on a rigorous holdout set. The article is more valuable for its debugging lessons than the happy path.

• •Fine-tuned google/muril-base-cased for code-mixed Hindi/Hinglish/English grievance classification: 87.3% accuracy, 0.869 macro F1
• •Four real bugs documented: eval leakage, shared RNG state, keyword bleed, Kaggle disk quota crash
• •MuRIL's multilingual pretraining bridges lexical gaps that break TF-IDF on code-mixed text

Generated with AI, which can make mistakes.