tojpaj/personal-finance-routing-model

Personal-Finance Query Routing Model (MuRIL, Hindi/Punjabi)

MuRIL fine-tuned for native Hindi/Punjabi financial-query intent classification (loan_inquiry/insurance/fraud_report/savings_scheme/ other). AutoScientist Challenge Part 2, Personal-Finance category.

Competitive note

An active competitor (Rome-1/personal-finance-adaptive-autoscientist, D->A, +80% quality) exists in this category using a generic English recipe with no native-language angle. This submission differentiates on the native-language axis rather than competing on generic quality.

Results

Accuracy 75.0%, Macro-F1 0.214 (held-out 20% split, 32 rows).

Known limitation, disclosed honestly: built without ULCA registration to fit the deadline, using IndicCorpV2 general web text instead of genuine financial queries. other dominates (132/158 training rows); the four genuine finance intents have too few examples (0-13 each) for reliable classification. The fraud_report hard-bypass safety design this idea was built around is documented but not meaningfully testable on this round's thin data. See the source project's PART2_SUBMISSION.md.