added subgroups for other mmlu variants

109ed1c7 · lintangsutawika · 93a45962 · 109ed1c7 · 109ed1c7 · 109ed1c7
Commit 109ed1c7 authored Oct 17, 2023 by lintangsutawika
20 changed files
--- a/lm_eval/tasks/mmlu/flan_n_shot/mmlu_gen_management.yaml
+++ b/lm_eval/tasks/mmlu/flan_n_shot/mmlu_gen_management.yaml
--- a/lm_eval/tasks/mmlu/flan_n_shot/mmlu_gen_marketing.yaml
+++ b/lm_eval/tasks/mmlu/flan_n_shot/mmlu_gen_marketing.yaml
--- a/lm_eval/tasks/mmlu/flan_n_shot/mmlu_gen_medical_genetics.yaml
+++ b/lm_eval/tasks/mmlu/flan_n_shot/mmlu_gen_medical_genetics.yaml
--- a/lm_eval/tasks/mmlu/flan_n_shot/mmlu_gen_miscellaneous.yaml
+++ b/lm_eval/tasks/mmlu/flan_n_shot/mmlu_gen_miscellaneous.yaml
--- a/lm_eval/tasks/mmlu/flan_n_shot/mmlu_gen_moral_disputes.yaml
+++ b/lm_eval/tasks/mmlu/flan_n_shot/mmlu_gen_moral_disputes.yaml
--- a/lm_eval/tasks/mmlu/flan_n_shot/mmlu_gen_moral_scenarios.yaml
+++ b/lm_eval/tasks/mmlu/flan_n_shot/mmlu_gen_moral_scenarios.yaml
--- a/lm_eval/tasks/mmlu/flan_n_shot/mmlu_gen_nutrition.yaml
+++ b/lm_eval/tasks/mmlu/flan_n_shot/mmlu_gen_nutrition.yaml
--- a/lm_eval/tasks/mmlu/flan_n_shot/mmlu_gen_philosophy.yaml
+++ b/lm_eval/tasks/mmlu/flan_n_shot/mmlu_gen_philosophy.yaml
--- a/lm_eval/tasks/mmlu/flan_n_shot/mmlu_gen_prehistory.yaml
+++ b/lm_eval/tasks/mmlu/flan_n_shot/mmlu_gen_prehistory.yaml
--- a/lm_eval/tasks/mmlu/flan_n_shot/mmlu_gen_professional_accounting.yaml
+++ b/lm_eval/tasks/mmlu/flan_n_shot/mmlu_gen_professional_accounting.yaml
--- a/lm_eval/tasks/mmlu/flan_n_shot/mmlu_gen_professional_law.yaml
+++ b/lm_eval/tasks/mmlu/flan_n_shot/mmlu_gen_professional_law.yaml
--- a/lm_eval/tasks/mmlu/flan_n_shot/mmlu_gen_professional_medicine.yaml
+++ b/lm_eval/tasks/mmlu/flan_n_shot/mmlu_gen_professional_medicine.yaml
--- a/lm_eval/tasks/mmlu/flan_n_shot/mmlu_gen_professional_psychology.yaml
+++ b/lm_eval/tasks/mmlu/flan_n_shot/mmlu_gen_professional_psychology.yaml
--- a/lm_eval/tasks/mmlu/flan_n_shot/mmlu_gen_public_relations.yaml
+++ b/lm_eval/tasks/mmlu/flan_n_shot/mmlu_gen_public_relations.yaml
--- a/lm_eval/tasks/mmlu/flan_n_shot/mmlu_gen_security_studies.yaml
+++ b/lm_eval/tasks/mmlu/flan_n_shot/mmlu_gen_security_studies.yaml
--- a/lm_eval/tasks/mmlu/flan_n_shot/mmlu_gen_sociology.yaml
+++ b/lm_eval/tasks/mmlu/flan_n_shot/mmlu_gen_sociology.yaml
--- a/lm_eval/tasks/mmlu/flan_n_shot/mmlu_gen_us_foreign_policy.yaml
+++ b/lm_eval/tasks/mmlu/flan_n_shot/mmlu_gen_us_foreign_policy.yaml
--- a/lm_eval/tasks/mmlu/flan_n_shot/mmlu_gen_virology.yaml
+++ b/lm_eval/tasks/mmlu/flan_n_shot/mmlu_gen_virology.yaml
--- a/lm_eval/tasks/mmlu/flan_n_shot/mmlu_gen_world_religions.yaml
+++ b/lm_eval/tasks/mmlu/flan_n_shot/mmlu_gen_world_religions.yaml
--- a/lm_eval/tasks/mmlu/flan_n_shot/loglikelihood/_mmlu.yaml
+++ b/lm_eval/tasks/mmlu/flan_n_shot/loglikelihood/_mmlu.yaml