Knowledge Distillation During Mid-Training Favors Reasoning over Factual Recall
Not assessedPlan blockedMeasurementclaim_t1_13b_trkd

Token-Routing Knowledge Distillation (TRKD) downstream evaluation performance after mid-training with OLMo-2 13B Instruct teacher.

Source: source_paper:Table 1, Page 9

Reported and observed measurements

downstream_accuracy

t1_13b_trkd_gsm8k

Reported 47.8 percentage_points

Observed — percentage_points

downstream_accuracy

t1_13b_trkd_gsm_s

Reported 32.7 percentage_points

Observed — percentage_points

downstream_accuracy

t1_13b_trkd_gsmplus

Reported 28.4 percentage_points

Observed — percentage_points

downstream_accuracy

t1_13b_trkd_bbh

Reported 29.3 percentage_points

Observed — percentage_points

downstream_accuracy

t1_13b_trkd_drop

Reported 34 percentage_points

Observed — percentage_points

downstream_accuracy

t1_13b_trkd_math

Reported 5.6 percentage_points

Observed — percentage_points

downstream_accuracy

t1_13b_trkd_tqa

Reported 53.8 percentage_points

Observed — percentage_points

downstream_accuracy

t1_13b_trkd_nq

Reported 24.2 percentage_points

Observed — percentage_points

downstream_accuracy

t1_13b_trkd_sqa

Reported 8.3 percentage_points

Observed — percentage_points

downstream_accuracy

t1_13b_trkd_mmlu

Reported 45.7 percentage_points

Observed — percentage_points

downstream_accuracy

t1_13b_trkd_mmlu_p

Reported 15.4 percentage_points

Observed — percentage_points

downstream_accuracy

t1_13b_trkd_arc_c

Reported 53 percentage_points

Observed — percentage_points

downstream_accuracy

t1_13b_trkd_obqa

Reported 56.6 percentage_points

Observed — percentage_points

downstream_accuracy

t1_13b_trkd_wino

Reported 50.7 percentage_points

Observed — percentage_points

downstream_accuracy

t1_13b_trkd_agi

Reported 35.2 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.