Downstream task performance and group macro-averages for Switch Distillation under routing threshold sweep 7b_q20.
来源：source_paper:Table 4, Page 21
报告指标与观测值
downstream_accuracy
t4_7b_q20_gsm8k
论文报告 69.7 percentage_points
实际观测 — percentage_points
downstream_accuracy
t4_7b_q20_gsm_s
论文报告 55.3 percentage_points
实际观测 — percentage_points
downstream_accuracy
t4_7b_q20_gsmplus
论文报告 46.1 percentage_points
实际观测 — percentage_points
downstream_accuracy
t4_7b_q20_bbh
论文报告 32.8 percentage_points
实际观测 — percentage_points
downstream_accuracy
t4_7b_q20_drop
论文报告 49.6 percentage_points
实际观测 — percentage_points
downstream_accuracy
t4_7b_q20_math
论文报告 14.8 percentage_points
实际观测 — percentage_points
downstream_accuracy
t4_7b_q20_avg_reasoning
论文报告 44.7 percentage_points
实际观测 — percentage_points
downstream_accuracy
t4_7b_q20_tqa
论文报告 54.9 percentage_points
实际观测 — percentage_points
downstream_accuracy
t4_7b_q20_nq
论文报告 24.6 percentage_points
实际观测 — percentage_points
downstream_accuracy
t4_7b_q20_sqa
论文报告 8.4 percentage_points
实际观测 — percentage_points
downstream_accuracy
t4_7b_q20_avg_factual
论文报告 29.3 percentage_points
实际观测 — percentage_points
downstream_accuracy
t4_7b_q20_mmlu
论文报告 51.6 percentage_points
实际观测 — percentage_points
downstream_accuracy
t4_7b_q20_mmlu_p
论文报告 19.8 percentage_points
实际观测 — percentage_points
downstream_accuracy
t4_7b_q20_arc_c
论文报告 64.7 percentage_points
实际观测 — percentage_points
downstream_accuracy
t4_7b_q20_obqa
论文报告 64.2 percentage_points
实际观测 — percentage_points
downstream_accuracy
t4_7b_q20_wino
论文报告 53.8 percentage_points
实际观测 — percentage_points
downstream_accuracy
t4_7b_q20_agi
论文报告 41.5 percentage_points
实际观测 — percentage_points
downstream_accuracy
t4_7b_q20_avg_knowledge
论文报告 49.3 percentage_points
实际观测 — percentage_points
Assessment（0）
这条 Claim 暂无已发布的不可变 Assessment。
关联运行（0）
这条 Claim 暂无关联执行。