Knowledge Distillation During Mid-Training Favors Reasoning over Factual Recall
Not assessedPlan blockedMeasurementclaim_t8_sft_7b_trkd

Downstream task evaluation after post-training stage SFT for 7b_trkd.

Source: source_paper:Table 8, Page 33, After SFT

Reported and observed measurements

downstream_accuracy

t8_sft_7b_trkd_gsm8k

Reported 49.9 percentage_points

Observed — percentage_points

downstream_accuracy

t8_sft_7b_trkd_gsm_s

Reported 31.1 percentage_points

Observed — percentage_points

downstream_accuracy

t8_sft_7b_trkd_gsmplus

Reported 26.9 percentage_points

Observed — percentage_points

downstream_accuracy

t8_sft_7b_trkd_bbh

Reported 31.5 percentage_points

Observed — percentage_points

downstream_accuracy

t8_sft_7b_trkd_drop

Reported 37.1 percentage_points

Observed — percentage_points

downstream_accuracy

t8_sft_7b_trkd_math

Reported 7.6 percentage_points

Observed — percentage_points

downstream_accuracy

t8_sft_7b_trkd_tqa

Reported 53.7 percentage_points

Observed — percentage_points

downstream_accuracy

t8_sft_7b_trkd_nq

Reported 23.1 percentage_points

Observed — percentage_points

downstream_accuracy

t8_sft_7b_trkd_sqa

Reported 8.4 percentage_points

Observed — percentage_points

downstream_accuracy

t8_sft_7b_trkd_mmlu

Reported 46.5 percentage_points

Observed — percentage_points

downstream_accuracy

t8_sft_7b_trkd_mmlu_p

Reported 17.4 percentage_points

Observed — percentage_points

downstream_accuracy

t8_sft_7b_trkd_arc_c

Reported 57.8 percentage_points

Observed — percentage_points

downstream_accuracy

t8_sft_7b_trkd_obqa

Reported 56 percentage_points

Observed — percentage_points

downstream_accuracy

t8_sft_7b_trkd_wino

Reported 51.4 percentage_points

Observed — percentage_points

downstream_accuracy

t8_sft_7b_trkd_agi

Reported 37.5 percentage_points

Observed — percentage_points

downstream_accuracy

t8_sft_7b_trkd_ife

Reported 45.3 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.