Natural Language Understanding on Downstream Benchmarks (HellaSwag, SIQA, PIQA, OBQA, WinoGrande, RACE) benchmark leaderboard · SOTA2 Research

Natural Language Understanding on Downstream Benchmarks (HellaSwag, SIQA, PIQA, OBQA, WinoGrande, RACE)

42.96HellaSwagFlashMHF

Evaluation Results

Method	Links
FlashMHF2025.12		42.96	44.17	68.44	27.8	54.46	22.26	43.35
FlashMHF2025.12		41.8	42.32	68.88	27.6	53.35	22.01	42.66
FlashMHF2025.12		39.8	42.17	67.08	27.2	51.3	22.62	41.7
Baseline2025.12		39.47	41.76	67.3	27.6	52.64	21.73	41.75
FlashMHF2025.12		33.97	40.63	66.27	27.8	52.41	21.8	40.48
FlashMHF2025.12		33.6	39.71	64.85	27.3	52.17	21.94	39.92
MH-FFN2025.12		33.45	41.3	65.18	27.6	51.46	21.87	40.14
Baseline2025.12		33.2	40.94	64.53	27.5	51.7	21.66	39.92
FlashMHF2025.12		32.57	41.5	65.34	27.6	53.35	21.59	40.32
Parametric KV2025.12		27.85	39.61	60.5	27.2	51.78	21.67	38.1