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_fkd

Forward KL Distillation (FKD, 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_fkd_gsm8k

论文报告 57.8 percentage_points

实际观测 — percentage_points

downstream_accuracy

t1_7b_fkd_gsm_s

论文报告 42.6 percentage_points

实际观测 — percentage_points

downstream_accuracy

t1_7b_fkd_gsmplus

论文报告 36.2 percentage_points

实际观测 — percentage_points

downstream_accuracy

t1_7b_fkd_bbh

论文报告 30.6 percentage_points

实际观测 — percentage_points

downstream_accuracy

t1_7b_fkd_drop

论文报告 41.9 percentage_points

实际观测 — percentage_points

downstream_accuracy

t1_7b_fkd_math

论文报告 7.6 percentage_points

实际观测 — percentage_points

downstream_accuracy

t1_7b_fkd_tqa

论文报告 53.9 percentage_points

实际观测 — percentage_points

downstream_accuracy

t1_7b_fkd_nq

论文报告 24.6 percentage_points

实际观测 — percentage_points

downstream_accuracy

t1_7b_fkd_sqa

论文报告 8.4 percentage_points

实际观测 — percentage_points

downstream_accuracy

t1_7b_fkd_mmlu

论文报告 49.5 percentage_points

实际观测 — percentage_points

downstream_accuracy

t1_7b_fkd_mmlu_p

论文报告 18.7 percentage_points

实际观测 — percentage_points

downstream_accuracy

t1_7b_fkd_arc_c

论文报告 61.9 percentage_points

实际观测 — percentage_points

downstream_accuracy

t1_7b_fkd_obqa

论文报告 61.8 percentage_points

实际观测 — percentage_points

downstream_accuracy

t1_7b_fkd_wino

论文报告 51.5 percentage_points

实际观测 — percentage_points

downstream_accuracy

t1_7b_fkd_agi

论文报告 38.6 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。