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_sd

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

来源:source_paper:Table 2, Page 9

报告指标与观测值

downstream_accuracy

t2_13b_sd_gsm8k

论文报告 77.8 percentage_points

实际观测 — percentage_points

downstream_accuracy

t2_13b_sd_gsm_s

论文报告 62.8 percentage_points

实际观测 — percentage_points

downstream_accuracy

t2_13b_sd_gsmplus

论文报告 51.8 percentage_points

实际观测 — percentage_points

downstream_accuracy

t2_13b_sd_bbh

论文报告 33.3 percentage_points

实际观测 — percentage_points

downstream_accuracy

t2_13b_sd_drop

论文报告 42.8 percentage_points

实际观测 — percentage_points

downstream_accuracy

t2_13b_sd_math

论文报告 19.6 percentage_points

实际观测 — percentage_points

downstream_accuracy

t2_13b_sd_tqa

论文报告 54.2 percentage_points

实际观测 — percentage_points

downstream_accuracy

t2_13b_sd_nq

论文报告 24.1 percentage_points

实际观测 — percentage_points

downstream_accuracy

t2_13b_sd_sqa

论文报告 8.4 percentage_points

实际观测 — percentage_points

downstream_accuracy

t2_13b_sd_mmlu

论文报告 43.1 percentage_points

实际观测 — percentage_points

downstream_accuracy

t2_13b_sd_mmlu_p

论文报告 18.1 percentage_points

实际观测 — percentage_points

downstream_accuracy

t2_13b_sd_arc_c

论文报告 56.5 percentage_points

实际观测 — percentage_points

downstream_accuracy

t2_13b_sd_obqa

论文报告 59 percentage_points

实际观测 — percentage_points

downstream_accuracy

t2_13b_sd_wino

论文报告 52.5 percentage_points

实际观测 — percentage_points

downstream_accuracy

t2_13b_sd_agi

论文报告 38.1 percentage_points

实际观测 — percentage_points

downstream_accuracy

t2_13b_sd_ife

论文报告 67.1 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。