merged main

lm_eval/tasks/kmmlu/hard/kmmlu_hard_machine_design_and_manufacturing.yaml

+dataset_name: machine_design_and_manufacturing
+include: _hard_kmmlu_yaml
+task: kmmlu_hard_machine_design_and_manufacturing

lm_eval/tasks/kmmlu/hard/kmmlu_hard_management.yaml

+dataset_name: management
+include: _hard_kmmlu_yaml
+task: kmmlu_hard_management

lm_eval/tasks/kmmlu/hard/kmmlu_hard_maritime_engineering.yaml

+dataset_name: maritime_engineering
+include: _hard_kmmlu_yaml
+task: kmmlu_hard_maritime_engineering

lm_eval/tasks/kmmlu/hard/kmmlu_hard_marketing.yaml

+dataset_name: marketing
+include: _hard_kmmlu_yaml
+task: kmmlu_hard_marketing

lm_eval/tasks/kmmlu/hard/kmmlu_hard_materials_engineering.yaml

+dataset_name: materials_engineering
+include: _hard_kmmlu_yaml
+task: kmmlu_hard_materials_engineering

lm_eval/tasks/kmmlu/hard/kmmlu_hard_math.yaml

+dataset_name: math
+include: _hard_kmmlu_yaml
+task: kmmlu_hard_math

lm_eval/tasks/kmmlu/hard/kmmlu_hard_mechanical_engineering.yaml

+dataset_name: mechanical_engineering
+include: _hard_kmmlu_yaml
+task: kmmlu_hard_mechanical_engineering

lm_eval/tasks/kmmlu/hard/kmmlu_hard_nondestructive_testing.yaml

+dataset_name: nondestructive_testing
+include: _hard_kmmlu_yaml
+task: kmmlu_hard_nondestructive_testing

lm_eval/tasks/kmmlu/hard/kmmlu_hard_patent.yaml

+dataset_name: patent
+include: _hard_kmmlu_yaml
+task: kmmlu_hard_patent

lm_eval/tasks/kmmlu/hard/kmmlu_hard_political_science_and_sociology.yaml

+dataset_name: political_science_and_sociology
+include: _hard_kmmlu_yaml
+task: kmmlu_hard_political_science_and_sociology

lm_eval/tasks/kmmlu/hard/kmmlu_hard_psychology.yaml

+dataset_name: psychology
+include: _hard_kmmlu_yaml
+task: kmmlu_hard_psychology

lm_eval/tasks/kmmlu/hard/kmmlu_hard_public_safety.yaml

+dataset_name: public_safety
+include: _hard_kmmlu_yaml
+task: kmmlu_hard_public_safety

lm_eval/tasks/kmmlu/hard/kmmlu_hard_railway_and_automotive_engineering.yaml

+dataset_name: railway_and_automotive_engineering
+include: _hard_kmmlu_yaml
+task: kmmlu_hard_railway_and_automotive_engineering

lm_eval/tasks/kmmlu/hard/kmmlu_hard_real_estate.yaml

+dataset_name: real_estate
+include: _hard_kmmlu_yaml
+task: kmmlu_hard_real_estate

lm_eval/tasks/kmmlu/hard/kmmlu_hard_refrigerating_machinery.yaml

+dataset_name: refrigerating_machinery
+include: _hard_kmmlu_yaml
+task: kmmlu_hard_refrigerating_machinery

lm_eval/tasks/kmmlu/hard/kmmlu_hard_social_welfare.yaml

+dataset_name: social_welfare
+include: _hard_kmmlu_yaml
+task: kmmlu_hard_social_welfare

lm_eval/tasks/kmmlu/hard/kmmlu_hard_taxation.yaml

+dataset_name: taxation
+include: _hard_kmmlu_yaml
+task: kmmlu_hard_taxation

lm_eval/tasks/kmmlu/hard/kmmlu_hard_telecommunications_and_wireless_technology.yaml

+dataset_name: telecommunications_and_wireless_technology
+include: _hard_kmmlu_yaml
+task: kmmlu_hard_telecommunications_and_wireless_technology

lm_eval/tasks/kmmlu/kmmlu_accounting.yaml

-"dataset_name": "Accounting"
-"include": "_default_kmmlu_yaml"
-"task": "kmmlu_accounting"

lm_eval/tasks/kmmlu/kmmlu_agricultural_sciences.yaml

-"dataset_name": "Agricultural-Sciences"
-"include": "_default_kmmlu_yaml"
-"task": "kmmlu_agricultural_sciences"