Upstream COMI-LINGUA options were 4-16 stringified per-word LID annotations that answered a different question than the natural-language instruction and were unlearnable in that shape. Rebuilt from state text via Unicode script presence. Same 3,220 rows; 3-class MCQ; gold distribution 2046/899/275.
fix: reshape hinglish_lid to 3-class script-detection MCQ (v1.1.0)
Upstream COMI-LINGUA options were 4-16 stringified per-word LID annotations that answered a different question than the natural-language instruction and were unlearnable in that shape. Rebuilt from state text via Unicode script presence. Same 3,220 rows; 3-class MCQ; gold distribution 2046/899/275.
26 days ago
Update leaderboard: Qwen3-4B and GPT-4o-mini results (2026-09-22)
