Cleaned BIRD-Platinum Dataset Propels Kimi-K2.6 Past Frontier Models in Text-to-SQL

Bridgewater AIA Labs and UIUC released BIRD-Platinum, a noise-corrected Text-to-SQL dataset; Kimi-K2.6 trained on it beat all tested frontier models at human-level accuracy and lower cost. The work also shows noisy data seriously degrades RLVR training, overturning prior robustness claims.

2026-08-28 ~ 2026-08-28 · 3 related posts

• Study: Noisy data severely degrades RLVR training performance — ddkang · 2026-08-28
• Study claims Text-to-SQL model achieves human-level accuracy at lower cost — ddkang · 2026-08-28
• Kimi-K2.6 trained on BIRD-Platinum outperforms frontier models — ddkang · 2026-08-28