Knowledge Distillation During Mid-Training Favors Reasoning over Factual Recall
尚未评估计划受阻指标结果claim_t9_random_routing

Per-task downstream accuracy for mid-training ablation random_routing using OLMo-2 7B Instruct teacher.

来源:source_paper:Table 9, Page 33

报告指标与观测值

downstream_accuracy

t9_random_routing_gsm8k

论文报告 61.5 percentage_points

实际观测 — percentage_points

downstream_accuracy

t9_random_routing_gsm_s

论文报告 45.5 percentage_points

实际观测 — percentage_points

downstream_accuracy

t9_random_routing_gsmplus

论文报告 38.6 percentage_points

实际观测 — percentage_points

downstream_accuracy

t9_random_routing_bbh

论文报告 32 percentage_points

实际观测 — percentage_points

downstream_accuracy

t9_random_routing_drop

论文报告 40.6 percentage_points

实际观测 — percentage_points

downstream_accuracy

t9_random_routing_math

论文报告 10.8 percentage_points

实际观测 — percentage_points

downstream_accuracy

t9_random_routing_tqa

论文报告 53.2 percentage_points

实际观测 — percentage_points

downstream_accuracy

t9_random_routing_nq

论文报告 23.9 percentage_points

实际观测 — percentage_points

downstream_accuracy

t9_random_routing_sqa

论文报告 8.5 percentage_points

实际观测 — percentage_points

downstream_accuracy

t9_random_routing_mmlu

论文报告 49.6 percentage_points

实际观测 — percentage_points

downstream_accuracy

t9_random_routing_mmlu_p

论文报告 17.8 percentage_points

实际观测 — percentage_points

downstream_accuracy

t9_random_routing_arc_c

论文报告 61.6 percentage_points

实际观测 — percentage_points

downstream_accuracy

t9_random_routing_obqa

论文报告 62.6 percentage_points

实际观测 — percentage_points

downstream_accuracy

t9_random_routing_wino

论文报告 52.6 percentage_points

实际观测 — percentage_points

downstream_accuracy

t9_random_routing_agi

论文报告 39.5 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。