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_rkd

Reverse KL Distillation (RKD, alpha=0.5) 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_rkd_gsm8k

论文报告 62.1 percentage_points

实际观测 — percentage_points

downstream_accuracy

t1_7b_rkd_gsm_s

论文报告 46.2 percentage_points

实际观测 — percentage_points

downstream_accuracy

t1_7b_rkd_gsmplus

论文报告 39.2 percentage_points

实际观测 — percentage_points

downstream_accuracy

t1_7b_rkd_bbh

论文报告 31.6 percentage_points

实际观测 — percentage_points

downstream_accuracy

t1_7b_rkd_drop

论文报告 43.1 percentage_points

实际观测 — percentage_points

downstream_accuracy

t1_7b_rkd_math

论文报告 10.4 percentage_points

实际观测 — percentage_points

downstream_accuracy

t1_7b_rkd_tqa

论文报告 53.1 percentage_points

实际观测 — percentage_points

downstream_accuracy

t1_7b_rkd_nq

论文报告 24.1 percentage_points

实际观测 — percentage_points

downstream_accuracy

t1_7b_rkd_sqa

论文报告 8.4 percentage_points

实际观测 — percentage_points

downstream_accuracy

t1_7b_rkd_mmlu

论文报告 50.1 percentage_points

实际观测 — percentage_points

downstream_accuracy

t1_7b_rkd_mmlu_p

论文报告 18.9 percentage_points

实际观测 — percentage_points

downstream_accuracy

t1_7b_rkd_arc_c

论文报告 62.5 percentage_points

实际观测 — percentage_points

downstream_accuracy

t1_7b_rkd_obqa

论文报告 62.8 percentage_points

实际观测 — percentage_points

downstream_accuracy

t1_7b_rkd_wino

论文报告 52.5 percentage_points

实际观测 — percentage_points

downstream_accuracy

t1_7b_rkd_agi

论文报告 39.3 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。