MuRIL fine-tuned for native Hindi/Punjabi workplace grievance intent classification (wage_dispute/harassment_escalation/leave_pf_esi/ general). AutoScientist Challenge Part 2, HR category.

Accuracy 82.1%, Macro-F1 0.225 (held-out 20% split, 28 rows).

Known limitation, disclosed honestly: built without ULCA registration to fit the deadline, using IndicCorpV2 general web text instead of genuine worker grievances. general dominates (127/137 training rows); the three genuine grievance intents have too few examples (1-5 each) for reliable classification. The escalation-bypass safety design this idea was built around (harassment_escalation + urgency=immediate -> hard-route to a human) is documented but not meaningfully testable on this round's thin urgency-label data (only 4 total). See the source project's PART2_SUBMISSION.md.