Knowledge Distillation During Mid-Training Favors Reasoning over Factual Recall
尚未评估计划受阻指标结果claim_t9_oracle_domain

Per-task downstream accuracy for mid-training ablation oracle_domain using OLMo-2 7B Instruct teacher.

来源:source_paper:Table 9, Page 33

报告指标与观测值

downstream_accuracy

t9_oracle_domain_gsm8k

论文报告 61 percentage_points

实际观测 — percentage_points

downstream_accuracy

t9_oracle_domain_gsm_s

论文报告 45.6 percentage_points

实际观测 — percentage_points

downstream_accuracy

t9_oracle_domain_gsmplus

论文报告 38.5 percentage_points

实际观测 — percentage_points

downstream_accuracy

t9_oracle_domain_bbh

论文报告 31.6 percentage_points

实际观测 — percentage_points

downstream_accuracy

t9_oracle_domain_drop

论文报告 40.1 percentage_points

实际观测 — percentage_points

downstream_accuracy

t9_oracle_domain_math

论文报告 8 percentage_points

实际观测 — percentage_points

downstream_accuracy

t9_oracle_domain_tqa

论文报告 53 percentage_points

实际观测 — percentage_points

downstream_accuracy

t9_oracle_domain_nq

论文报告 22.6 percentage_points

实际观测 — percentage_points

downstream_accuracy

t9_oracle_domain_sqa

论文报告 8.3 percentage_points

实际观测 — percentage_points

downstream_accuracy

t9_oracle_domain_mmlu

论文报告 49.4 percentage_points

实际观测 — percentage_points

downstream_accuracy

t9_oracle_domain_mmlu_p

论文报告 18.8 percentage_points

实际观测 — percentage_points

downstream_accuracy

t9_oracle_domain_arc_c

论文报告 61.1 percentage_points

实际观测 — percentage_points

downstream_accuracy

t9_oracle_domain_obqa

论文报告 61.4 percentage_points

实际观测 — percentage_points

downstream_accuracy

t9_oracle_domain_wino

论文报告 52.6 percentage_points

实际观测 — percentage_points

downstream_accuracy

t9_oracle_domain_agi

论文报告 38.9 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。