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_trkd

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

来源:source_paper:Table 2, Page 9

报告指标与观测值

downstream_accuracy

t2_7b_trkd_gsm8k

论文报告 70.7 percentage_points

实际观测 — percentage_points

downstream_accuracy

t2_7b_trkd_gsm_s

论文报告 52.9 percentage_points

实际观测 — percentage_points

downstream_accuracy

t2_7b_trkd_gsmplus

论文报告 46.3 percentage_points

实际观测 — percentage_points

downstream_accuracy

t2_7b_trkd_bbh

论文报告 31.8 percentage_points

实际观测 — percentage_points

downstream_accuracy

t2_7b_trkd_drop

论文报告 36.8 percentage_points

实际观测 — percentage_points

downstream_accuracy

t2_7b_trkd_math

论文报告 14.4 percentage_points

实际观测 — percentage_points

downstream_accuracy

t2_7b_trkd_tqa

论文报告 51.7 percentage_points

实际观测 — percentage_points

downstream_accuracy

t2_7b_trkd_nq

论文报告 22.3 percentage_points

实际观测 — percentage_points

downstream_accuracy

t2_7b_trkd_sqa

论文报告 8 percentage_points

实际观测 — percentage_points

downstream_accuracy

t2_7b_trkd_mmlu

论文报告 40.9 percentage_points

实际观测 — percentage_points

downstream_accuracy

t2_7b_trkd_mmlu_p

论文报告 17.4 percentage_points

实际观测 — percentage_points

downstream_accuracy

t2_7b_trkd_arc_c

论文报告 59 percentage_points

实际观测 — percentage_points

downstream_accuracy

t2_7b_trkd_obqa

论文报告 53.6 percentage_points

实际观测 — percentage_points

downstream_accuracy

t2_7b_trkd_wino

论文报告 51.5 percentage_points

实际观测 — percentage_points

downstream_accuracy

t2_7b_trkd_agi

论文报告 35.6 percentage_points

实际观测 — percentage_points

downstream_accuracy

t2_7b_trkd_ife

论文报告 65.2 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。