Knowledge Distillation During Mid-Training Favors Reasoning over Factual Recall
尚未评估计划受阻指标结果claim_t2_13b_rkd

Full downstream results after standard 4-stage post-training pipeline (SFT, DPO, RLVR1, RLVR2) for 13B RKD post-training performance.

来源:source_paper:Table 2, Page 9

报告指标与观测值

downstream_accuracy

t2_13b_rkd_gsm8k

论文报告 75.1 percentage_points

实际观测 — percentage_points

downstream_accuracy

t2_13b_rkd_gsm_s

论文报告 58.1 percentage_points

实际观测 — percentage_points

downstream_accuracy

t2_13b_rkd_gsmplus

论文报告 49.3 percentage_points

实际观测 — percentage_points

downstream_accuracy

t2_13b_rkd_bbh

论文报告 33.3 percentage_points

实际观测 — percentage_points

downstream_accuracy

t2_13b_rkd_drop

论文报告 38.7 percentage_points

实际观测 — percentage_points

downstream_accuracy

t2_13b_rkd_math

论文报告 16 percentage_points

实际观测 — percentage_points

downstream_accuracy

t2_13b_rkd_tqa

论文报告 53.9 percentage_points

实际观测 — percentage_points

downstream_accuracy

t2_13b_rkd_nq

论文报告 22.3 percentage_points

实际观测 — percentage_points

downstream_accuracy

t2_13b_rkd_sqa

论文报告 7.7 percentage_points

实际观测 — percentage_points

downstream_accuracy

t2_13b_rkd_mmlu

论文报告 38.2 percentage_points

实际观测 — percentage_points

downstream_accuracy

t2_13b_rkd_mmlu_p

论文报告 17 percentage_points

实际观测 — percentage_points

downstream_accuracy

t2_13b_rkd_arc_c

论文报告 56 percentage_points

实际观测 — percentage_points

downstream_accuracy

t2_13b_rkd_obqa

论文报告 55.2 percentage_points

实际观测 — percentage_points

downstream_accuracy

t2_13b_rkd_wino

论文报告 51.1 percentage_points

实际观测 — percentage_points

downstream_accuracy

t2_13b_rkd_agi

论文报告 36.6 percentage_points

实际观测 — percentage_points

downstream_accuracy

t2_13b_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。