Knowledge Distillation During Mid-Training Favors Reasoning over Factual Recall
尚未评估计划受阻指标结果claim_t9_always_ce

Per-task downstream accuracy for mid-training ablation always_ce 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_always_ce_gsm8k

论文报告 69.1 percentage_points

实际观测 — percentage_points

downstream_accuracy

t9_always_ce_gsm_s

论文报告 55.2 percentage_points

实际观测 — percentage_points

downstream_accuracy

t9_always_ce_gsmplus

论文报告 46.5 percentage_points

实际观测 — percentage_points

downstream_accuracy

t9_always_ce_bbh

论文报告 33.3 percentage_points

实际观测 — percentage_points

downstream_accuracy

t9_always_ce_drop

论文报告 50.1 percentage_points

实际观测 — percentage_points

downstream_accuracy

t9_always_ce_math

论文报告 12 percentage_points

实际观测 — percentage_points

downstream_accuracy

t9_always_ce_tqa

论文报告 54.3 percentage_points

实际观测 — percentage_points

downstream_accuracy

t9_always_ce_nq

论文报告 24.5 percentage_points

实际观测 — percentage_points

downstream_accuracy

t9_always_ce_sqa

论文报告 9.4 percentage_points

实际观测 — percentage_points

downstream_accuracy

t9_always_ce_mmlu

论文报告 51.1 percentage_points

实际观测 — percentage_points

downstream_accuracy

t9_always_ce_mmlu_p

论文报告 19.9 percentage_points

实际观测 — percentage_points

downstream_accuracy

t9_always_ce_arc_c

论文报告 63.7 percentage_points

实际观测 — percentage_points

downstream_accuracy

t9_always_ce_obqa

论文报告 63.6 percentage_points

实际观测 — percentage_points

downstream_accuracy

t9_always_ce_wino

论文报告 52.7 percentage_points

实际观测 — percentage_points

downstream_accuracy

t9_always_ce_agi

论文报告 41.1 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。