Downstream task performance and group macro-averages for Switch Distillation under routing threshold sweep 13b_q20.
Source: source_paper:Table 4, Page 21
Reported and observed measurements
downstream_accuracy
t4_13b_q20_gsm8k
Reported 66 percentage_points
Observed — percentage_points
downstream_accuracy
t4_13b_q20_gsm_s
Reported 54.8 percentage_points
Observed — percentage_points
downstream_accuracy
t4_13b_q20_gsmplus
Reported 42.2 percentage_points
Observed — percentage_points
downstream_accuracy
t4_13b_q20_bbh
Reported 31.6 percentage_points
Observed — percentage_points
downstream_accuracy
t4_13b_q20_drop
Reported 45.8 percentage_points
Observed — percentage_points
downstream_accuracy
t4_13b_q20_math
Reported 12 percentage_points
Observed — percentage_points
downstream_accuracy
t4_13b_q20_avg_reasoning
Reported 42.1 percentage_points
Observed — percentage_points
downstream_accuracy
t4_13b_q20_tqa
Reported 54.1 percentage_points
Observed — percentage_points
downstream_accuracy
t4_13b_q20_nq
Reported 24.9 percentage_points
Observed — percentage_points
downstream_accuracy
t4_13b_q20_sqa
Reported 9 percentage_points
Observed — percentage_points
downstream_accuracy
t4_13b_q20_avg_factual
Reported 29.3 percentage_points
Observed — percentage_points
downstream_accuracy
t4_13b_q20_mmlu
Reported 48.5 percentage_points
Observed — percentage_points
downstream_accuracy
t4_13b_q20_mmlu_p
Reported 17.7 percentage_points
Observed — percentage_points
downstream_accuracy
t4_13b_q20_arc_c
Reported 58.5 percentage_points
Observed — percentage_points
downstream_accuracy
t4_13b_q20_obqa
Reported 63.8 percentage_points
Observed — percentage_points
downstream_accuracy
t4_13b_q20_wino
Reported 51.2 percentage_points
Observed — percentage_points
downstream_accuracy
t4_13b_q20_agi
Reported 39.2 percentage_points
Observed — percentage_points
downstream_accuracy
t4_13b_q20_avg_knowledge
Reported 46.5 percentage_points
Observed — percentage_points
Assessments (0)
No immutable Assessment has been published for this Claim yet.
Runs (0)
No execution has been linked to this Claim yet.