Knowledge Distillation During Mid-Training Favors Reasoning over Factual Recall
Not assessedPlan blockedMeasurementclaim_t1_7b_trkd

Token-Routing Knowledge Distillation (TRKD) downstream evaluation performance after mid-training with OLMo-2 7B Instruct teacher.

Source: source_paper:Table 1, Page 9

Reported and observed measurements

downstream_accuracy

t1_7b_trkd_gsm8k

Reported 52.5 percentage_points

Observed — percentage_points

downstream_accuracy

t1_7b_trkd_gsm_s

Reported 36.9 percentage_points

Observed — percentage_points

downstream_accuracy

t1_7b_trkd_gsmplus

Reported 31 percentage_points

Observed — percentage_points

downstream_accuracy

t1_7b_trkd_bbh

Reported 30.3 percentage_points

Observed — percentage_points

downstream_accuracy

t1_7b_trkd_drop

Reported 36.8 percentage_points

Observed — percentage_points

downstream_accuracy

t1_7b_trkd_math

Reported 5.6 percentage_points

Observed — percentage_points

downstream_accuracy

t1_7b_trkd_tqa

Reported 54 percentage_points

Observed — percentage_points

downstream_accuracy

t1_7b_trkd_nq

Reported 24.1 percentage_points

Observed — percentage_points

downstream_accuracy

t1_7b_trkd_sqa

Reported 7.8 percentage_points

Observed — percentage_points

downstream_accuracy

t1_7b_trkd_mmlu

Reported 47.9 percentage_points

Observed — percentage_points

downstream_accuracy

t1_7b_trkd_mmlu_p

Reported 17.4 percentage_points

Observed — percentage_points

downstream_accuracy

t1_7b_trkd_arc_c

Reported 58.9 percentage_points

Observed — percentage_points

downstream_accuracy

t1_7b_trkd_obqa

Reported 60 percentage_points

Observed — percentage_points

downstream_accuracy

t1_7b_trkd_wino

Reported 51.2 percentage_points

Observed — percentage_points

downstream_accuracy

t1_7b_trkd_agi

Reported 37.5 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.