Knowledge Distillation During Mid-Training Favors Reasoning over Factual Recall
尚未评估计划受阻指标结果claim_t1_7b_trkd

Token-Routing Knowledge Distillation (TRKD) downstream evaluation performance after mid-training with OLMo-2 7B Instruct teacher.

来源:source_paper:Table 1, Page 9

报告指标与观测值

downstream_accuracy

t1_7b_trkd_gsm8k

论文报告 52.5 percentage_points

实际观测 — percentage_points

downstream_accuracy

t1_7b_trkd_gsm_s

论文报告 36.9 percentage_points

实际观测 — percentage_points

downstream_accuracy

t1_7b_trkd_gsmplus

论文报告 31 percentage_points

实际观测 — percentage_points

downstream_accuracy

t1_7b_trkd_bbh

论文报告 30.3 percentage_points

实际观测 — percentage_points

downstream_accuracy

t1_7b_trkd_drop

论文报告 36.8 percentage_points

实际观测 — percentage_points

downstream_accuracy

t1_7b_trkd_math

论文报告 5.6 percentage_points

实际观测 — percentage_points

downstream_accuracy

t1_7b_trkd_tqa

论文报告 54 percentage_points

实际观测 — percentage_points

downstream_accuracy

t1_7b_trkd_nq

论文报告 24.1 percentage_points

实际观测 — percentage_points

downstream_accuracy

t1_7b_trkd_sqa

论文报告 7.8 percentage_points

实际观测 — percentage_points

downstream_accuracy

t1_7b_trkd_mmlu

论文报告 47.9 percentage_points

实际观测 — percentage_points

downstream_accuracy

t1_7b_trkd_mmlu_p

论文报告 17.4 percentage_points

实际观测 — percentage_points

downstream_accuracy

t1_7b_trkd_arc_c

论文报告 58.9 percentage_points

实际观测 — percentage_points

downstream_accuracy

t1_7b_trkd_obqa

论文报告 60 percentage_points

实际观测 — percentage_points

downstream_accuracy

t1_7b_trkd_wino

论文报告 51.2 percentage_points

实际观测 — percentage_points

downstream_accuracy

t1_7b_trkd_agi

论文报告 37.5 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。