ResearchBenchmarksCommonsense Reasoning on HellaSwag ArabicFollow39.25Accuracy (Normalized)ML35.786836.685937.58538.4841Apr 22, 2026Evaluation ResultsMethodMethodLinksAccuracy (Normalized)ML2026.0439.25No Filtering2026.0439.09HQ seedVariant=SeedVariant=Seed2026.0439.09MKC-e2026.0438.82HQ2026.0438.73ML (Q3)Sampling Strategy=Q3Sampling Strategy=Q32026.0438.7ML (20%)Retention=20%Retention=20%2026.0438.57ML (10%)Retention=10%Retention=10%2026.0437.47HQ (20%)Retention=20%Retention=20%2026.0437.28HQ (10%)Retention=10%Retention=10%2026.0435.92