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_trkd

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

来源:source_paper:Table 2, Page 9

报告指标与观测值

downstream_accuracy

t2_13b_trkd_gsm8k

论文报告 69.6 percentage_points

实际观测 — percentage_points

downstream_accuracy

t2_13b_trkd_gsm_s

论文报告 48.5 percentage_points

实际观测 — percentage_points

downstream_accuracy

t2_13b_trkd_gsmplus

论文报告 42.6 percentage_points

实际观测 — percentage_points

downstream_accuracy

t2_13b_trkd_bbh

论文报告 31.5 percentage_points

实际观测 — percentage_points

downstream_accuracy

t2_13b_trkd_drop

论文报告 34.6 percentage_points

实际观测 — percentage_points

downstream_accuracy

t2_13b_trkd_math

论文报告 12.2 percentage_points

实际观测 — percentage_points

downstream_accuracy

t2_13b_trkd_tqa

论文报告 53.9 percentage_points

实际观测 — percentage_points

downstream_accuracy

t2_13b_trkd_nq

论文报告 22.2 percentage_points

实际观测 — percentage_points

downstream_accuracy

t2_13b_trkd_sqa

论文报告 8.2 percentage_points

实际观测 — percentage_points

downstream_accuracy

t2_13b_trkd_mmlu

论文报告 37 percentage_points

实际观测 — percentage_points

downstream_accuracy

t2_13b_trkd_mmlu_p

论文报告 17.3 percentage_points

实际观测 — percentage_points

downstream_accuracy

t2_13b_trkd_arc_c

论文报告 55.4 percentage_points

实际观测 — percentage_points

downstream_accuracy

t2_13b_trkd_obqa

论文报告 54 percentage_points

实际观测 — percentage_points

downstream_accuracy

t2_13b_trkd_wino

论文报告 51.3 percentage_points

实际观测 — percentage_points

downstream_accuracy

t2_13b_trkd_agi

论文报告 33.6 percentage_points

实际观测 — percentage_points

downstream_accuracy

t2_13b_trkd_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。