switch mmlu variants over to using

22b95671 · haileyschoelkopf · c1d9e625 · 22b95671 · 22b95671 · 22b95671
Commit 22b95671 authored Jul 02, 2024 by haileyschoelkopf
20 changed files
--- a/lm_eval/tasks/mmlu/flan_n_shot/loglikelihood/mmlu_college_physics.yaml
+++ b/lm_eval/tasks/mmlu/flan_n_shot/loglikelihood/mmlu_college_physics.yaml
--- a/lm_eval/tasks/mmlu/flan_n_shot/loglikelihood/mmlu_computer_security.yaml
+++ b/lm_eval/tasks/mmlu/flan_n_shot/loglikelihood/mmlu_computer_security.yaml
--- a/lm_eval/tasks/mmlu/flan_n_shot/loglikelihood/mmlu_conceptual_physics.yaml
+++ b/lm_eval/tasks/mmlu/flan_n_shot/loglikelihood/mmlu_conceptual_physics.yaml
--- a/lm_eval/tasks/mmlu/flan_n_shot/loglikelihood/mmlu_econometrics.yaml
+++ b/lm_eval/tasks/mmlu/flan_n_shot/loglikelihood/mmlu_econometrics.yaml
--- a/lm_eval/tasks/mmlu/flan_n_shot/loglikelihood/mmlu_electrical_engineering.yaml
+++ b/lm_eval/tasks/mmlu/flan_n_shot/loglikelihood/mmlu_electrical_engineering.yaml
--- a/lm_eval/tasks/mmlu/flan_n_shot/loglikelihood/mmlu_elementary_mathematics.yaml
+++ b/lm_eval/tasks/mmlu/flan_n_shot/loglikelihood/mmlu_elementary_mathematics.yaml
--- a/lm_eval/tasks/mmlu/flan_n_shot/loglikelihood/mmlu_formal_logic.yaml
+++ b/lm_eval/tasks/mmlu/flan_n_shot/loglikelihood/mmlu_formal_logic.yaml
--- a/lm_eval/tasks/mmlu/flan_n_shot/loglikelihood/mmlu_global_facts.yaml
+++ b/lm_eval/tasks/mmlu/flan_n_shot/loglikelihood/mmlu_global_facts.yaml
--- a/lm_eval/tasks/mmlu/flan_n_shot/loglikelihood/mmlu_high_school_biology.yaml
+++ b/lm_eval/tasks/mmlu/flan_n_shot/loglikelihood/mmlu_high_school_biology.yaml
--- a/lm_eval/tasks/mmlu/flan_n_shot/loglikelihood/mmlu_high_school_chemistry.yaml
+++ b/lm_eval/tasks/mmlu/flan_n_shot/loglikelihood/mmlu_high_school_chemistry.yaml
--- a/lm_eval/tasks/mmlu/flan_n_shot/loglikelihood/mmlu_high_school_computer_science.yaml
+++ b/lm_eval/tasks/mmlu/flan_n_shot/loglikelihood/mmlu_high_school_computer_science.yaml
--- a/lm_eval/tasks/mmlu/flan_n_shot/loglikelihood/mmlu_high_school_european_history.yaml
+++ b/lm_eval/tasks/mmlu/flan_n_shot/loglikelihood/mmlu_high_school_european_history.yaml
--- a/lm_eval/tasks/mmlu/flan_n_shot/loglikelihood/mmlu_high_school_geography.yaml
+++ b/lm_eval/tasks/mmlu/flan_n_shot/loglikelihood/mmlu_high_school_geography.yaml
--- a/lm_eval/tasks/mmlu/flan_n_shot/loglikelihood/mmlu_high_school_government_and_politics.yaml
+++ b/lm_eval/tasks/mmlu/flan_n_shot/loglikelihood/mmlu_high_school_government_and_politics.yaml
--- a/lm_eval/tasks/mmlu/flan_n_shot/loglikelihood/mmlu_high_school_macroeconomics.yaml
+++ b/lm_eval/tasks/mmlu/flan_n_shot/loglikelihood/mmlu_high_school_macroeconomics.yaml
--- a/lm_eval/tasks/mmlu/flan_n_shot/loglikelihood/mmlu_high_school_mathematics.yaml
+++ b/lm_eval/tasks/mmlu/flan_n_shot/loglikelihood/mmlu_high_school_mathematics.yaml
--- a/lm_eval/tasks/mmlu/flan_n_shot/loglikelihood/mmlu_high_school_microeconomics.yaml
+++ b/lm_eval/tasks/mmlu/flan_n_shot/loglikelihood/mmlu_high_school_microeconomics.yaml
--- a/lm_eval/tasks/mmlu/flan_n_shot/loglikelihood/mmlu_high_school_physics.yaml
+++ b/lm_eval/tasks/mmlu/flan_n_shot/loglikelihood/mmlu_high_school_physics.yaml
--- a/lm_eval/tasks/mmlu/flan_n_shot/loglikelihood/mmlu_high_school_psychology.yaml
+++ b/lm_eval/tasks/mmlu/flan_n_shot/loglikelihood/mmlu_high_school_psychology.yaml
--- a/lm_eval/tasks/mmlu/flan_n_shot/loglikelihood/mmlu_high_school_statistics.yaml
+++ b/lm_eval/tasks/mmlu/flan_n_shot/loglikelihood/mmlu_high_school_statistics.yaml