Knowledge Distillation During Mid-Training Favors Reasoning over Factual Recall
尚未评估计划受阻指标结果claim_t4_13b_q10

Downstream task performance and group macro-averages for Switch Distillation under routing threshold sweep 13b_q10.

来源:source_paper:Table 4, Page 21

报告指标与观测值

downstream_accuracy

t4_13b_q10_gsm8k

论文报告 62.1 percentage_points

实际观测 — percentage_points

downstream_accuracy

t4_13b_q10_gsm_s

论文报告 49.6 percentage_points

实际观测 — percentage_points

downstream_accuracy

t4_13b_q10_gsmplus

论文报告 38.3 percentage_points

实际观测 — percentage_points

downstream_accuracy

t4_13b_q10_bbh

论文报告 28.6 percentage_points

实际观测 — percentage_points

downstream_accuracy

t4_13b_q10_drop

论文报告 44.2 percentage_points

实际观测 — percentage_points

downstream_accuracy

t4_13b_q10_math

论文报告 7.8 percentage_points

实际观测 — percentage_points

downstream_accuracy

t4_13b_q10_avg_reasoning

论文报告 38.5 percentage_points

实际观测 — percentage_points

downstream_accuracy

t4_13b_q10_tqa

论文报告 52.2 percentage_points

实际观测 — percentage_points

downstream_accuracy

t4_13b_q10_nq

论文报告 23.1 percentage_points

实际观测 — percentage_points

downstream_accuracy

t4_13b_q10_sqa

论文报告 9.2 percentage_points

实际观测 — percentage_points

downstream_accuracy

t4_13b_q10_avg_factual

论文报告 28.1 percentage_points

实际观测 — percentage_points

downstream_accuracy

t4_13b_q10_mmlu

论文报告 46.8 percentage_points

实际观测 — percentage_points

downstream_accuracy

t4_13b_q10_mmlu_p

论文报告 17.1 percentage_points

实际观测 — percentage_points

downstream_accuracy

t4_13b_q10_arc_c

论文报告 55.5 percentage_points

实际观测 — percentage_points

downstream_accuracy

t4_13b_q10_obqa

论文报告 59.6 percentage_points

实际观测 — percentage_points

downstream_accuracy

t4_13b_q10_wino

论文报告 52.7 percentage_points

实际观测 — percentage_points

downstream_accuracy

t4_13b_q10_agi

论文报告 37.9 percentage_points

实际观测 — percentage_points

downstream_accuracy

t4_13b_q10_avg_knowledge

论文报告 45 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。