ResearchBenchmarksMultitask Language Understanding on M-MMLU-c (acc_norm)Multitask Language Understanding on M-MMLU-c (acc_norm)Follow26.74Normalized Accuracy (M-MMLU-c)HQ (20%)26.11626.27826.4426.602Apr 22, 2026Evaluation resultsMethodMethodLinksNormalized Accuracy (M-MMLU-c)HQ (20%)Retention=20%Retention=20%2026.0426.74No Filtering2026.0426.69HQ seedVariant=SeedVariant=Seed2026.0426.53HQ (10%)Retention=10%Retention=10%2026.0426.52ML2026.0426.46ML (20%)Retention=20%Retention=20%2026.0426.34MKC-e2026.0426.28ML (10%)Retention=10%Retention=10%2026.0426.18ML (Q3)Sampling Strategy=Q3Sampling Strategy=Q32026.0426.15HQ2026.0426.14