Knowledge Distillation During Mid-Training Favors Reasoning over Factual Recall
尚未评估计划受阻指标结果claim_t9_teacher_top1

Per-task downstream accuracy for mid-training ablation teacher_top1 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_teacher_top1_gsm8k

论文报告 62.2 percentage_points

实际观测 — percentage_points

downstream_accuracy

t9_teacher_top1_gsm_s

论文报告 46 percentage_points

实际观测 — percentage_points

downstream_accuracy

t9_teacher_top1_gsmplus

论文报告 40.1 percentage_points

实际观测 — percentage_points

downstream_accuracy

t9_teacher_top1_bbh

论文报告 30.7 percentage_points

实际观测 — percentage_points

downstream_accuracy

t9_teacher_top1_drop

论文报告 42.1 percentage_points

实际观测 — percentage_points

downstream_accuracy

t9_teacher_top1_math

论文报告 8.8 percentage_points

实际观测 — percentage_points

downstream_accuracy

t9_teacher_top1_tqa

论文报告 57.9 percentage_points

实际观测 — percentage_points

downstream_accuracy

t9_teacher_top1_nq

论文报告 24.7 percentage_points

实际观测 — percentage_points

downstream_accuracy

t9_teacher_top1_sqa

论文报告 9.3 percentage_points

实际观测 — percentage_points

downstream_accuracy

t9_teacher_top1_mmlu

论文报告 48.8 percentage_points

实际观测 — percentage_points

downstream_accuracy

t9_teacher_top1_mmlu_p

论文报告 18 percentage_points

实际观测 — percentage_points

downstream_accuracy

t9_teacher_top1_arc_c

论文报告 59.8 percentage_points

实际观测 — percentage_points

downstream_accuracy

t9_teacher_top1_obqa

论文报告 62 percentage_points

实际观测 — percentage_points

downstream_accuracy

t9_teacher_top1_wino

论文报告 51 percentage_points

实际观测 — percentage_points

downstream_accuracy

t9_teacher_top1_agi

论文报告 39.2 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。