Knowledge Distillation During Mid-Training Favors Reasoning over Factual Recall
尚未评估计划受阻指标结果claim_t2_7b_fkd

Full downstream results after standard 4-stage post-training pipeline (SFT, DPO, RLVR1, RLVR2) for 7B FKD post-training performance.

来源:source_paper:Table 2, Page 9

报告指标与观测值

downstream_accuracy

t2_7b_fkd_gsm8k

论文报告 76.2 percentage_points

实际观测 — percentage_points

downstream_accuracy

t2_7b_fkd_gsm_s

论文报告 56.4 percentage_points

实际观测 — percentage_points

downstream_accuracy

t2_7b_fkd_gsmplus

论文报告 51.6 percentage_points

实际观测 — percentage_points

downstream_accuracy

t2_7b_fkd_bbh

论文报告 33.8 percentage_points

实际观测 — percentage_points

downstream_accuracy

t2_7b_fkd_drop

论文报告 38.4 percentage_points

实际观测 — percentage_points

downstream_accuracy

t2_7b_fkd_math

论文报告 18 percentage_points

实际观测 — percentage_points

downstream_accuracy

t2_7b_fkd_tqa

论文报告 51.9 percentage_points

实际观测 — percentage_points

downstream_accuracy

t2_7b_fkd_nq

论文报告 22.2 percentage_points

实际观测 — percentage_points

downstream_accuracy

t2_7b_fkd_sqa

论文报告 8.1 percentage_points

实际观测 — percentage_points

downstream_accuracy

t2_7b_fkd_mmlu

论文报告 42.2 percentage_points

实际观测 — percentage_points

downstream_accuracy

t2_7b_fkd_mmlu_p

论文报告 18.6 percentage_points

实际观测 — percentage_points

downstream_accuracy

t2_7b_fkd_arc_c

论文报告 60.2 percentage_points

实际观测 — percentage_points

downstream_accuracy

t2_7b_fkd_obqa

论文报告 57.4 percentage_points

实际观测 — percentage_points

downstream_accuracy

t2_7b_fkd_wino

论文报告 51.5 percentage_points

实际观测 — percentage_points

downstream_accuracy

t2_7b_fkd_agi

论文报告 38 percentage_points

实际观测 — percentage_points

downstream_accuracy

t2_7b_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。