tojpaj/agriculture-routing-model
Agriculture Query Routing Model (MuRIL, Hindi/Punjabi)
MuRIL fine-tuned for native Hindi/Punjabi farmer-query intent classification (pestid/irrigation/mandiprice/scheme_eligibility/ other). AutoScientist Challenge Part 2, Agriculture category.
Results
Accuracy 92.5%, Macro-F1 0.240 (held-out 20% split, 40 rows).
Known limitation, disclosed honestly — the most extreme in this round's five submissions: built without ULCA registration (multi-day external dependency) to fit the deadline, using IndicCorpV2 general web text instead of genuine farmer queries. The high accuracy is almost entirely the other class's dominance (182/200 training rows), not genuine skill distinguishing pestid/irrigation/mandiprice/ schemeeligibility (2-6 examples each). The pipeline is validated and ready for real ULCA farmer-query data; see the source project's `PART2SUBMISSION.md` for the full writeup.
