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_ntp_shared

Downstream task evaluation after post-training stage DPO for ntp_shared.

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

报告指标与观测值

downstream_accuracy

t8_dpo_ntp_shared_gsm8k

论文报告 52.4 percentage_points

实际观测 — percentage_points

downstream_accuracy

t8_dpo_ntp_shared_gsm_s

论文报告 33.4 percentage_points

实际观测 — percentage_points

downstream_accuracy

t8_dpo_ntp_shared_gsmplus

论文报告 28.2 percentage_points

实际观测 — percentage_points

downstream_accuracy

t8_dpo_ntp_shared_bbh

论文报告 32.3 percentage_points

实际观测 — percentage_points

downstream_accuracy

t8_dpo_ntp_shared_drop

论文报告 33.5 percentage_points

实际观测 — percentage_points

downstream_accuracy

t8_dpo_ntp_shared_math

论文报告 6.6 percentage_points

实际观测 — percentage_points

downstream_accuracy

t8_dpo_ntp_shared_tqa

论文报告 55.4 percentage_points

实际观测 — percentage_points

downstream_accuracy

t8_dpo_ntp_shared_nq

论文报告 21.5 percentage_points

实际观测 — percentage_points

downstream_accuracy

t8_dpo_ntp_shared_sqa

论文报告 7.7 percentage_points

实际观测 — percentage_points

downstream_accuracy

t8_dpo_ntp_shared_mmlu

论文报告 41.1 percentage_points

实际观测 — percentage_points

downstream_accuracy

t8_dpo_ntp_shared_mmlu_p

论文报告 16.4 percentage_points

实际观测 — percentage_points

downstream_accuracy

t8_dpo_ntp_shared_arc_c

论文报告 50.3 percentage_points

实际观测 — percentage_points

downstream_accuracy

t8_dpo_ntp_shared_obqa

论文报告 50.6 percentage_points

实际观测 — percentage_points

downstream_accuracy

t8_dpo_ntp_shared_wino

论文报告 51.5 percentage_points

实际观测 — percentage_points

downstream_accuracy

t8_dpo_ntp_shared_agi

论文报告 34.1 percentage_points

实际观测 — percentage_points

downstream_accuracy

t8_dpo_ntp_shared_ife

论文报告 59.9 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。