Knowledge Distillation During Mid-Training Favors Reasoning over Factual Recall
Not assessedPlan blockedMeasurementclaim_t4_13b_q30

Downstream task performance and group macro-averages for Switch Distillation under routing threshold sweep 13b_q30.

Source: source_paper:Table 4, Page 21

Reported and observed measurements

downstream_accuracy

t4_13b_q30_gsm8k

Reported 66 percentage_points

Observed — percentage_points

downstream_accuracy

t4_13b_q30_gsm_s

Reported 53.1 percentage_points

Observed — percentage_points

downstream_accuracy

t4_13b_q30_gsmplus

Reported 42.9 percentage_points

Observed — percentage_points

downstream_accuracy

t4_13b_q30_bbh

Reported 32.5 percentage_points

Observed — percentage_points

downstream_accuracy

t4_13b_q30_drop

Reported 44.4 percentage_points

Observed — percentage_points

downstream_accuracy

t4_13b_q30_math

Reported 11 percentage_points

Observed — percentage_points

downstream_accuracy

t4_13b_q30_avg_reasoning

Reported 41.6 percentage_points

Observed — percentage_points

downstream_accuracy

t4_13b_q30_tqa

Reported 54.2 percentage_points

Observed — percentage_points

downstream_accuracy

t4_13b_q30_nq

Reported 24.7 percentage_points

Observed — percentage_points

downstream_accuracy

t4_13b_q30_sqa

Reported 8.6 percentage_points

Observed — percentage_points

downstream_accuracy

t4_13b_q30_avg_factual

Reported 29.2 percentage_points

Observed — percentage_points

downstream_accuracy

t4_13b_q30_mmlu

Reported 49 percentage_points

Observed — percentage_points

downstream_accuracy

t4_13b_q30_mmlu_p

Reported 18.4 percentage_points

Observed — percentage_points

downstream_accuracy

t4_13b_q30_arc_c

Reported 60.9 percentage_points

Observed — percentage_points

downstream_accuracy

t4_13b_q30_obqa

Reported 63.6 percentage_points

Observed — percentage_points

downstream_accuracy

t4_13b_q30_wino

Reported 52.9 percentage_points

Observed — percentage_points

downstream_accuracy

t4_13b_q30_agi

Reported 38.7 percentage_points

Observed — percentage_points

downstream_accuracy

t4_13b_q30_avg_knowledge

Reported 47.2 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.