Scheme-Anchor-Check-Scheme-Router

Purpose of Dataset

This model predicts which government scheme a citizen's question is about — currently PM-KISAN or Ayushman Bharat PM-JAY — given the question text in Hinglish. It's the routing component a real scheme-assistance system would need before it can even look up the correct guideline document to ground its answer in. The model is a TF-IDF vectorizer (unigrams and bigrams) feeding a Logistic Regression classifier, trained on the companion Scheme-Anchor-Check QA Dataset (14 questions across 2 schemes). This model is intended to be used together with the Scheme-Anchor-Check evaluation pipeline: once a question is routed to the correct scheme, a separate grounded-QA process (using the actual verified scheme guideline as context) answers the question and gets scored on faithfulness to the source document — this model handles the routing step of that pipeline, not the answering or faithfulness-scoring steps. Limitation: trained on only 2 schemes and 14 examples — accuracy and usefulness will improve significantly as more real, verified government schemes are added to the training set.