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_trkd

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

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

报告指标与观测值

downstream_accuracy

t8_dpo_7b_trkd_gsm8k

论文报告 57.2 percentage_points

实际观测 — percentage_points

downstream_accuracy

t8_dpo_7b_trkd_gsm_s

论文报告 34.9 percentage_points

实际观测 — percentage_points

downstream_accuracy

t8_dpo_7b_trkd_gsmplus

论文报告 32.3 percentage_points

实际观测 — percentage_points

downstream_accuracy

t8_dpo_7b_trkd_bbh

论文报告 32.4 percentage_points

实际观测 — percentage_points

downstream_accuracy

t8_dpo_7b_trkd_drop

论文报告 36.6 percentage_points

实际观测 — percentage_points

downstream_accuracy

t8_dpo_7b_trkd_math

论文报告 7.2 percentage_points

实际观测 — percentage_points

downstream_accuracy

t8_dpo_7b_trkd_tqa

论文报告 53.1 percentage_points

实际观测 — percentage_points

downstream_accuracy

t8_dpo_7b_trkd_nq

论文报告 23.2 percentage_points

实际观测 — percentage_points

downstream_accuracy

t8_dpo_7b_trkd_sqa

论文报告 7.9 percentage_points

实际观测 — percentage_points

downstream_accuracy

t8_dpo_7b_trkd_mmlu

论文报告 47 percentage_points

实际观测 — percentage_points

downstream_accuracy

t8_dpo_7b_trkd_mmlu_p

论文报告 17.9 percentage_points

实际观测 — percentage_points

downstream_accuracy

t8_dpo_7b_trkd_arc_c

论文报告 56.7 percentage_points

实际观测 — percentage_points

downstream_accuracy

t8_dpo_7b_trkd_obqa

论文报告 57.2 percentage_points

实际观测 — percentage_points

downstream_accuracy

t8_dpo_7b_trkd_wino

论文报告 51.9 percentage_points

实际观测 — percentage_points

downstream_accuracy

t8_dpo_7b_trkd_agi

论文报告 37.5 percentage_points

实际观测 — percentage_points

downstream_accuracy

t8_dpo_7b_trkd_ife

论文报告 62.5 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。