Reverse KL Distillation (RKD, alpha=0.5) downstream evaluation performance after mid-training with OLMo-2 13B Instruct teacher.
来源:source_paper:Table 1, Page 9
报告指标与观测值
downstream_accuracy
t1_13b_rkd_gsm8k
论文报告 59.2 percentage_points
实际观测 — percentage_points
downstream_accuracy
t1_13b_rkd_gsm_s
论文报告 48.7 percentage_points
实际观测 — percentage_points
downstream_accuracy
t1_13b_rkd_gsmplus
论文报告 37.4 percentage_points
实际观测 — percentage_points
downstream_accuracy
t1_13b_rkd_bbh
论文报告 31.4 percentage_points
实际观测 — percentage_points
downstream_accuracy
t1_13b_rkd_drop
论文报告 37.3 percentage_points
实际观测 — percentage_points
downstream_accuracy
t1_13b_rkd_math
论文报告 8 percentage_points
实际观测 — percentage_points
downstream_accuracy
t1_13b_rkd_tqa
论文报告 53.4 percentage_points
实际观测 — percentage_points
downstream_accuracy
t1_13b_rkd_nq
论文报告 24 percentage_points
实际观测 — percentage_points
downstream_accuracy
t1_13b_rkd_sqa
论文报告 8.5 percentage_points
实际观测 — percentage_points
downstream_accuracy
t1_13b_rkd_mmlu
论文报告 48.4 percentage_points
实际观测 — percentage_points
downstream_accuracy
t1_13b_rkd_mmlu_p
论文报告 17.3 percentage_points
实际观测 — percentage_points
downstream_accuracy
t1_13b_rkd_arc_c
论文报告 57.8 percentage_points
实际观测 — percentage_points
downstream_accuracy
t1_13b_rkd_obqa
论文报告 60 percentage_points
实际观测 — percentage_points
downstream_accuracy
t1_13b_rkd_wino
论文报告 51 percentage_points
实际观测 — percentage_points
downstream_accuracy
t1_13b_rkd_agi
论文报告 38.3 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。