Standard Next-Token Prediction (NTP) baseline downstream evaluation performance after 60B mid-training tokens across 15 tasks covering Reasoning, Factual Recall, and Knowledge & Commonsense.
来源：source_paper:Table 1, Page 9
报告指标与观测值
downstream_accuracy
t1_ntp_gsm8k
论文报告 40.4 percentage_points
实际观测 — percentage_points
downstream_accuracy
t1_ntp_gsm_s
论文报告 29.8 percentage_points
实际观测 — percentage_points
downstream_accuracy
t1_ntp_gsmplus
论文报告 23.1 percentage_points
实际观测 — percentage_points
downstream_accuracy
t1_ntp_bbh
论文报告 29.9 percentage_points
实际观测 — percentage_points
downstream_accuracy
t1_ntp_drop
论文报告 29.8 percentage_points
实际观测 — percentage_points
downstream_accuracy
t1_ntp_math
论文报告 3.8 percentage_points
实际观测 — percentage_points
downstream_accuracy
t1_ntp_tqa
论文报告 56.7 percentage_points
实际观测 — percentage_points
downstream_accuracy
t1_ntp_nq
论文报告 25.5 percentage_points
实际观测 — percentage_points
downstream_accuracy
t1_ntp_sqa
论文报告 8.7 percentage_points
实际观测 — percentage_points
downstream_accuracy
t1_ntp_mmlu
论文报告 43.6 percentage_points
实际观测 — percentage_points
downstream_accuracy
t1_ntp_mmlu_p
论文报告 15.5 percentage_points
实际观测 — percentage_points
downstream_accuracy
t1_ntp_arc_c
论文报告 51.1 percentage_points
实际观测 — percentage_points
downstream_accuracy
t1_ntp_obqa
论文报告 51.8 percentage_points
实际观测 — percentage_points
downstream_accuracy
t1_ntp_wino
论文报告 51.4 percentage_points
实际观测 — percentage_points
downstream_accuracy
t1_ntp_agi
论文报告 34.1 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。