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_13b_fkd

Downstream task evaluation after post-training stage DPO for 13b_fkd.

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

报告指标与观测值

downstream_accuracy

t8_dpo_13b_fkd_gsm8k

论文报告 61 percentage_points

实际观测 — percentage_points

downstream_accuracy

t8_dpo_13b_fkd_gsm_s

论文报告 34.2 percentage_points

实际观测 — percentage_points

downstream_accuracy

t8_dpo_13b_fkd_gsmplus

论文报告 31.7 percentage_points

实际观测 — percentage_points

downstream_accuracy

t8_dpo_13b_fkd_bbh

论文报告 33.4 percentage_points

实际观测 — percentage_points

downstream_accuracy

t8_dpo_13b_fkd_drop

论文报告 36.9 percentage_points

实际观测 — percentage_points

downstream_accuracy

t8_dpo_13b_fkd_math

论文报告 8.2 percentage_points

实际观测 — percentage_points

downstream_accuracy

t8_dpo_13b_fkd_tqa

论文报告 54.6 percentage_points

实际观测 — percentage_points

downstream_accuracy

t8_dpo_13b_fkd_nq

论文报告 23.2 percentage_points

实际观测 — percentage_points

downstream_accuracy

t8_dpo_13b_fkd_sqa

论文报告 7.7 percentage_points

实际观测 — percentage_points

downstream_accuracy

t8_dpo_13b_fkd_mmlu

论文报告 45.9 percentage_points

实际观测 — percentage_points

downstream_accuracy

t8_dpo_13b_fkd_mmlu_p

论文报告 17.4 percentage_points

实际观测 — percentage_points

downstream_accuracy

t8_dpo_13b_fkd_arc_c

论文报告 53.2 percentage_points

实际观测 — percentage_points

downstream_accuracy

t8_dpo_13b_fkd_obqa

论文报告 57.2 percentage_points

实际观测 — percentage_points

downstream_accuracy

t8_dpo_13b_fkd_wino

论文报告 52.3 percentage_points

实际观测 — percentage_points

downstream_accuracy

t8_dpo_13b_fkd_agi

论文报告 37.4 percentage_points

实际观测 — percentage_points

downstream_accuracy

t8_dpo_13b_fkd_ife

论文报告 64.7 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。