Knowledge Distillation During Mid-Training Favors Reasoning over Factual Recall
尚未评估计划受阻指标结果claim_t1_7b_sd

Switch Distillation (SD, q=20%, tau=2) downstream evaluation performance after mid-training with OLMo-2 7B Instruct teacher, demonstrating substantial gains on reasoning while maintaining factual recall.

来源:source_paper:Table 1, Page 9

报告指标与观测值

downstream_accuracy

t1_7b_sd_gsm8k

论文报告 69.7 percentage_points

实际观测 — percentage_points

downstream_accuracy

t1_7b_sd_gsm_s

论文报告 55.3 percentage_points

实际观测 — percentage_points

downstream_accuracy

t1_7b_sd_gsmplus

论文报告 46.1 percentage_points

实际观测 — percentage_points

downstream_accuracy

t1_7b_sd_bbh

论文报告 32.8 percentage_points

实际观测 — percentage_points

downstream_accuracy

t1_7b_sd_drop

论文报告 49.6 percentage_points

实际观测 — percentage_points

downstream_accuracy

t1_7b_sd_math

论文报告 14.8 percentage_points

实际观测 — percentage_points

downstream_accuracy

t1_7b_sd_tqa

论文报告 54.9 percentage_points

实际观测 — percentage_points

downstream_accuracy

t1_7b_sd_nq

论文报告 24.6 percentage_points

实际观测 — percentage_points

downstream_accuracy

t1_7b_sd_sqa

论文报告 8.4 percentage_points

实际观测 — percentage_points

downstream_accuracy

t1_7b_sd_mmlu

论文报告 51.6 percentage_points

实际观测 — percentage_points

downstream_accuracy

t1_7b_sd_mmlu_p

论文报告 19.8 percentage_points

实际观测 — percentage_points

downstream_accuracy

t1_7b_sd_arc_c

论文报告 64.7 percentage_points

实际观测 — percentage_points

downstream_accuracy

t1_7b_sd_obqa

论文报告 64.2 percentage_points

实际观测 — percentage_points

downstream_accuracy

t1_7b_sd_wino

论文报告 53.8 percentage_points

实际观测 — percentage_points

downstream_accuracy

t1_7b_sd_agi

论文报告 41.5 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。