Knowledge Distillation During Mid-Training Favors Reasoning over Factual Recall
尚未评估计划受阻指标结果claim_t4_7b_q10

Downstream task performance and group macro-averages for Switch Distillation under routing threshold sweep 7b_q10.

来源:source_paper:Table 4, Page 21

报告指标与观测值

downstream_accuracy

t4_7b_q10_gsm8k

论文报告 61.6 percentage_points

实际观测 — percentage_points

downstream_accuracy

t4_7b_q10_gsm_s

论文报告 48.8 percentage_points

实际观测 — percentage_points

downstream_accuracy

t4_7b_q10_gsmplus

论文报告 39.4 percentage_points

实际观测 — percentage_points

downstream_accuracy

t4_7b_q10_bbh

论文报告 32.5 percentage_points

实际观测 — percentage_points

downstream_accuracy

t4_7b_q10_drop

论文报告 50.2 percentage_points

实际观测 — percentage_points

downstream_accuracy

t4_7b_q10_math

论文报告 11.4 percentage_points

实际观测 — percentage_points

downstream_accuracy

t4_7b_q10_avg_reasoning

论文报告 40.6 percentage_points

实际观测 — percentage_points

downstream_accuracy

t4_7b_q10_tqa

论文报告 53.7 percentage_points

实际观测 — percentage_points

downstream_accuracy

t4_7b_q10_nq

论文报告 23.9 percentage_points

实际观测 — percentage_points

downstream_accuracy

t4_7b_q10_sqa

论文报告 8.6 percentage_points

实际观测 — percentage_points

downstream_accuracy

t4_7b_q10_avg_factual

论文报告 28.7 percentage_points

实际观测 — percentage_points

downstream_accuracy

t4_7b_q10_mmlu

论文报告 50.6 percentage_points

实际观测 — percentage_points

downstream_accuracy

t4_7b_q10_mmlu_p

论文报告 19 percentage_points

实际观测 — percentage_points

downstream_accuracy

t4_7b_q10_arc_c

论文报告 62 percentage_points

实际观测 — percentage_points

downstream_accuracy

t4_7b_q10_obqa

论文报告 63.8 percentage_points

实际观测 — percentage_points

downstream_accuracy

t4_7b_q10_wino

论文报告 55.5 percentage_points

实际观测 — percentage_points

downstream_accuracy

t4_7b_q10_agi

论文报告 40.4 percentage_points

实际观测 — percentage_points

downstream_accuracy

t4_7b_q10_avg_knowledge

论文报告 48.5 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。