Forward KL Distillation (FKD, 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_fkd_gsm8k
论文报告 52.7 percentage_points
实际观测 — percentage_points
downstream_accuracy
t1_13b_fkd_gsm_s
论文报告 37.9 percentage_points
实际观测 — percentage_points
downstream_accuracy
t1_13b_fkd_gsmplus
论文报告 31.8 percentage_points
实际观测 — percentage_points
downstream_accuracy
t1_13b_fkd_bbh
论文报告 29.4 percentage_points
实际观测 — percentage_points
downstream_accuracy
t1_13b_fkd_drop
论文报告 33.1 percentage_points
实际观测 — percentage_points
downstream_accuracy
t1_13b_fkd_math
论文报告 6.4 percentage_points
实际观测 — percentage_points
downstream_accuracy
t1_13b_fkd_tqa
论文报告 54.2 percentage_points
实际观测 — percentage_points
downstream_accuracy
t1_13b_fkd_nq
论文报告 24.7 percentage_points
实际观测 — percentage_points
downstream_accuracy
t1_13b_fkd_sqa
论文报告 8.1 percentage_points
实际观测 — percentage_points
downstream_accuracy
t1_13b_fkd_mmlu
论文报告 47.5 percentage_points
实际观测 — percentage_points
downstream_accuracy
t1_13b_fkd_mmlu_p
论文报告 16.3 percentage_points
实际观测 — percentage_points
downstream_accuracy
t1_13b_fkd_arc_c
论文报告 55.3 percentage_points
实际观测 — percentage_points
downstream_accuracy
t1_13b_fkd_obqa
论文报告 56.2 percentage_points
实际观测 — percentage_points
downstream_accuracy
t1_13b_fkd_wino
论文报告 51.2 percentage_points
实际观测 — percentage_points
downstream_accuracy
t1_13b_fkd_agi
论文报告 36.2 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。