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_correct

Per-task downstream accuracy for mid-training ablation teacher_correct 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_correct_gsm8k

论文报告 63.8 percentage_points

实际观测 — percentage_points

downstream_accuracy

t9_teacher_correct_gsm_s

论文报告 49.5 percentage_points

实际观测 — percentage_points

downstream_accuracy

t9_teacher_correct_gsmplus

论文报告 41.7 percentage_points

实际观测 — percentage_points

downstream_accuracy

t9_teacher_correct_bbh

论文报告 31.8 percentage_points

实际观测 — percentage_points

downstream_accuracy

t9_teacher_correct_drop

论文报告 45.2 percentage_points

实际观测 — percentage_points

downstream_accuracy

t9_teacher_correct_math

论文报告 9.8 percentage_points

实际观测 — percentage_points

downstream_accuracy

t9_teacher_correct_tqa

论文报告 44.5 percentage_points

实际观测 — percentage_points

downstream_accuracy

t9_teacher_correct_nq

论文报告 19.4 percentage_points

实际观测 — percentage_points

downstream_accuracy

t9_teacher_correct_sqa

论文报告 8.7 percentage_points

实际观测 — percentage_points

downstream_accuracy

t9_teacher_correct_mmlu

论文报告 50.6 percentage_points

实际观测 — percentage_points

downstream_accuracy

t9_teacher_correct_mmlu_p

论文报告 19.3 percentage_points

实际观测 — percentage_points

downstream_accuracy

t9_teacher_correct_arc_c

论文报告 63.2 percentage_points

实际观测 — percentage_points

downstream_accuracy

t9_teacher_correct_obqa

论文报告 64.6 percentage_points

实际观测 — percentage_points

downstream_accuracy

t9_teacher_correct_wino

论文报告 52 percentage_points

实际观测 — percentage_points

downstream_accuracy

t9_teacher_correct_agi

论文报告 39.9 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。