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_q30

Downstream task performance and group macro-averages for Switch Distillation under routing threshold sweep 7b_q30.

来源:source_paper:Table 4, Page 21

报告指标与观测值

downstream_accuracy

t4_7b_q30_gsm8k

论文报告 69.8 percentage_points

实际观测 — percentage_points

downstream_accuracy

t4_7b_q30_gsm_s

论文报告 54.1 percentage_points

实际观测 — percentage_points

downstream_accuracy

t4_7b_q30_gsmplus

论文报告 45.6 percentage_points

实际观测 — percentage_points

downstream_accuracy

t4_7b_q30_bbh

论文报告 33.8 percentage_points

实际观测 — percentage_points

downstream_accuracy

t4_7b_q30_drop

论文报告 47.2 percentage_points

实际观测 — percentage_points

downstream_accuracy

t4_7b_q30_math

论文报告 13 percentage_points

实际观测 — percentage_points

downstream_accuracy

t4_7b_q30_avg_reasoning

论文报告 43.9 percentage_points

实际观测 — percentage_points

downstream_accuracy

t4_7b_q30_tqa

论文报告 54.7 percentage_points

实际观测 — percentage_points

downstream_accuracy

t4_7b_q30_nq

论文报告 25.3 percentage_points

实际观测 — percentage_points

downstream_accuracy

t4_7b_q30_sqa

论文报告 8.8 percentage_points

实际观测 — percentage_points

downstream_accuracy

t4_7b_q30_avg_factual

论文报告 29.6 percentage_points

实际观测 — percentage_points

downstream_accuracy

t4_7b_q30_mmlu

论文报告 51 percentage_points

实际观测 — percentage_points

downstream_accuracy

t4_7b_q30_mmlu_p

论文报告 19.5 percentage_points

实际观测 — percentage_points

downstream_accuracy

t4_7b_q30_arc_c

论文报告 63.5 percentage_points

实际观测 — percentage_points

downstream_accuracy

t4_7b_q30_obqa

论文报告 62.4 percentage_points

实际观测 — percentage_points

downstream_accuracy

t4_7b_q30_wino

论文报告 53 percentage_points

实际观测 — percentage_points

downstream_accuracy

t4_7b_q30_agi

论文报告 40.7 percentage_points

实际观测 — percentage_points

downstream_accuracy

t4_7b_q30_avg_knowledge

论文报告 48.4 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。