Knowledge Distillation During Mid-Training Favors Reasoning over Factual Recall
尚未评估计划受阻指标结果claim_t9_sd_fkl

Per-task downstream accuracy for mid-training ablation sd_fkl 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_sd_fkl_gsm8k

论文报告 66.7 percentage_points

实际观测 — percentage_points

downstream_accuracy

t9_sd_fkl_gsm_s

论文报告 52 percentage_points

实际观测 — percentage_points

downstream_accuracy

t9_sd_fkl_gsmplus

论文报告 43.7 percentage_points

实际观测 — percentage_points

downstream_accuracy

t9_sd_fkl_bbh

论文报告 31.2 percentage_points

实际观测 — percentage_points

downstream_accuracy

t9_sd_fkl_drop

论文报告 46.4 percentage_points

实际观测 — percentage_points

downstream_accuracy

t9_sd_fkl_math

论文报告 11 percentage_points

实际观测 — percentage_points

downstream_accuracy

t9_sd_fkl_tqa

论文报告 54.8 percentage_points

实际观测 — percentage_points

downstream_accuracy

t9_sd_fkl_nq

论文报告 24.3 percentage_points

实际观测 — percentage_points

downstream_accuracy

t9_sd_fkl_sqa

论文报告 8.2 percentage_points

实际观测 — percentage_points

downstream_accuracy

t9_sd_fkl_mmlu

论文报告 50.5 percentage_points

实际观测 — percentage_points

downstream_accuracy

t9_sd_fkl_mmlu_p

论文报告 19.1 percentage_points

实际观测 — percentage_points

downstream_accuracy

t9_sd_fkl_arc_c

论文报告 63.7 percentage_points

实际观测 — percentage_points

downstream_accuracy

t9_sd_fkl_obqa

论文报告 62.6 percentage_points

实际观测 — percentage_points

downstream_accuracy

t9_sd_fkl_wino

论文报告 52.2 percentage_points

实际观测 — percentage_points

downstream_accuracy

t9_sd_fkl_agi

论文报告 39.6 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。