Knowledge Distillation During Mid-Training Favors Reasoning over Factual Recall
尚未评估计划受阻指标结果claim_t8_dpo_7b_rkd

Downstream task evaluation after post-training stage DPO for 7b_rkd.

来源:source_paper:Table 8, Page 33, After DPO

报告指标与观测值

downstream_accuracy

t8_dpo_7b_rkd_gsm8k

论文报告 64.5 percentage_points

实际观测 — percentage_points

downstream_accuracy

t8_dpo_7b_rkd_gsm_s

论文报告 40 percentage_points

实际观测 — percentage_points

downstream_accuracy

t8_dpo_7b_rkd_gsmplus

论文报告 36.4 percentage_points

实际观测 — percentage_points

downstream_accuracy

t8_dpo_7b_rkd_bbh

论文报告 34.2 percentage_points

实际观测 — percentage_points

downstream_accuracy

t8_dpo_7b_rkd_drop

论文报告 40.5 percentage_points

实际观测 — percentage_points

downstream_accuracy

t8_dpo_7b_rkd_math

论文报告 10 percentage_points

实际观测 — percentage_points

downstream_accuracy

t8_dpo_7b_rkd_tqa

论文报告 52.8 percentage_points

实际观测 — percentage_points

downstream_accuracy

t8_dpo_7b_rkd_nq

论文报告 22.3 percentage_points

实际观测 — percentage_points

downstream_accuracy

t8_dpo_7b_rkd_sqa

论文报告 7.8 percentage_points

实际观测 — percentage_points

downstream_accuracy

t8_dpo_7b_rkd_mmlu

论文报告 48.8 percentage_points

实际观测 — percentage_points

downstream_accuracy

t8_dpo_7b_rkd_mmlu_p

论文报告 17.3 percentage_points

实际观测 — percentage_points

downstream_accuracy

t8_dpo_7b_rkd_arc_c

论文报告 57.4 percentage_points

实际观测 — percentage_points

downstream_accuracy

t8_dpo_7b_rkd_obqa

论文报告 55 percentage_points

实际观测 — percentage_points

downstream_accuracy

t8_dpo_7b_rkd_wino

论文报告 52.5 percentage_points

实际观测 — percentage_points

downstream_accuracy

t8_dpo_7b_rkd_agi

论文报告 38.2 percentage_points

实际观测 — percentage_points

downstream_accuracy

t8_dpo_7b_rkd_ife

论文报告 64.1 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。