Multi-task Language Understanding on MergeBench Llama-3.2-3B (test) benchmark leaderboard · SOTA2 Research

Multi-task Language Understanding on MergeBench Llama-3.2-3B (test)

44.1Average ScoreSWUDI

Evaluation Results

Method	Links
SWUDI2026.06		44.1	58.45	37.8	57.67	24.58	45.02	47.01	37.55	44.69
TIES w/ DARE2026.06		43.56	52.99	35.37	57.41	25.51	47.42	47.55	37.04	45.19
SWUDI-A2026.06		43.46	58.15	36.59	56.35	22.18	45.42	46.85	37.3	44.83
OptMerge2026.06		43.37	53.53	34.76	58.99	25.51	45.23	46.99	37.3	44.66
TSV Merging2026.06		43.24	55.72	36.59	56.88	20.15	46.28	48.17	37.13	45.01
WUDI Merging2026.06		42.2	52.54	37.2	57.14	17.93	46.56	44.86	37.21	44.17
TA w/ DARE2026.06		42.01	46.7	33.54	59.26	18.67	47.8	48.12	36.44	45.56
Task Arithmetic2026.06		40.92	44.73	33.54	59.79	14.42	47.38	46.34	36.27	44.88
Iso-C2026.06		40.12	48.22	35.37	55.56	8.13	44.83	47.49	37.04	44.29
Weight Average2026.06		39.49	42.76	31.71	59.52	9.24	46.02	46.35	35.76	44.53
TIES Merging2026.06		39.11	42.61	30.49	57.14	7.58	44.91	48.25	37.13	44.75