Knowledge Distillation During Mid-Training Favors Reasoning over Factual Recall
尚未评估计划受阻指标结果claim_t8_sft_7b_sd

Downstream task evaluation after post-training stage SFT for 7b_sd.

来源:source_paper:Table 8, Page 33, After SFT

报告指标与观测值

downstream_accuracy

t8_sft_7b_sd_gsm8k

论文报告 63.7 percentage_points

实际观测 — percentage_points

downstream_accuracy

t8_sft_7b_sd_gsm_s

论文报告 42.7 percentage_points

实际观测 — percentage_points

downstream_accuracy

t8_sft_7b_sd_gsmplus

论文报告 36.7 percentage_points

实际观测 — percentage_points

downstream_accuracy

t8_sft_7b_sd_bbh

论文报告 33.5 percentage_points

实际观测 — percentage_points

downstream_accuracy

t8_sft_7b_sd_drop

论文报告 48.9 percentage_points

实际观测 — percentage_points

downstream_accuracy

t8_sft_7b_sd_math

论文报告 12.8 percentage_points

实际观测 — percentage_points

downstream_accuracy

t8_sft_7b_sd_tqa

论文报告 55.1 percentage_points

实际观测 — percentage_points

downstream_accuracy

t8_sft_7b_sd_nq

论文报告 24.5 percentage_points

实际观测 — percentage_points

downstream_accuracy

t8_sft_7b_sd_sqa

论文报告 8.5 percentage_points

实际观测 — percentage_points

downstream_accuracy

t8_sft_7b_sd_mmlu

论文报告 50.3 percentage_points

实际观测 — percentage_points

downstream_accuracy

t8_sft_7b_sd_mmlu_p

论文报告 19.4 percentage_points

实际观测 — percentage_points

downstream_accuracy

t8_sft_7b_sd_arc_c

论文报告 61.3 percentage_points

实际观测 — percentage_points

downstream_accuracy

t8_sft_7b_sd_obqa

论文报告 61.4 percentage_points

实际观测 — percentage_points

downstream_accuracy

t8_sft_7b_sd_wino

论文报告 56.1 percentage_points

实际观测 — percentage_points

downstream_accuracy

t8_sft_7b_sd_agi

论文报告 40.6 percentage_points

实际观测 — percentage_points

downstream_accuracy

t8_sft_7b_sd_ife

论文报告 45.5 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。