Downstream task performance and group macro-averages for Switch Distillation under routing threshold sweep 7b_q10.
Source: source_paper:Table 4, Page 21
Reported and observed measurements
downstream_accuracy
t4_7b_q10_gsm8k
Reported 61.6 percentage_points
Observed — percentage_points
downstream_accuracy
t4_7b_q10_gsm_s
Reported 48.8 percentage_points
Observed — percentage_points
downstream_accuracy
t4_7b_q10_gsmplus
Reported 39.4 percentage_points
Observed — percentage_points
downstream_accuracy
t4_7b_q10_bbh
Reported 32.5 percentage_points
Observed — percentage_points
downstream_accuracy
t4_7b_q10_drop
Reported 50.2 percentage_points
Observed — percentage_points
downstream_accuracy
t4_7b_q10_math
Reported 11.4 percentage_points
Observed — percentage_points
downstream_accuracy
t4_7b_q10_avg_reasoning
Reported 40.6 percentage_points
Observed — percentage_points
downstream_accuracy
t4_7b_q10_tqa
Reported 53.7 percentage_points
Observed — percentage_points
downstream_accuracy
t4_7b_q10_nq
Reported 23.9 percentage_points
Observed — percentage_points
downstream_accuracy
t4_7b_q10_sqa
Reported 8.6 percentage_points
Observed — percentage_points
downstream_accuracy
t4_7b_q10_avg_factual
Reported 28.7 percentage_points
Observed — percentage_points
downstream_accuracy
t4_7b_q10_mmlu
Reported 50.6 percentage_points
Observed — percentage_points
downstream_accuracy
t4_7b_q10_mmlu_p
Reported 19 percentage_points
Observed — percentage_points
downstream_accuracy
t4_7b_q10_arc_c
Reported 62 percentage_points
Observed — percentage_points
downstream_accuracy
t4_7b_q10_obqa
Reported 63.8 percentage_points
Observed — percentage_points
downstream_accuracy
t4_7b_q10_wino
Reported 55.5 percentage_points
Observed — percentage_points
downstream_accuracy
t4_7b_q10_agi
Reported 40.4 percentage_points
Observed — percentage_points
downstream_accuracy
t4_7b_q10_avg_knowledge
Reported 48.5 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.