Add files using upload-large-folder tool
Browse filesThis view is limited to 50 files because it contains too many changes. See raw diff
- lm-evaluation-harness/lm_eval/tasks/global_mmlu/full/mg/global_mmlu_full_mg_human_sexuality.yaml +5 -0
- lm-evaluation-harness/lm_eval/tasks/global_mmlu/full/mg/global_mmlu_full_mg_jurisprudence.yaml +5 -0
- lm-evaluation-harness/lm_eval/tasks/global_mmlu/full/mg/global_mmlu_full_mg_logical_fallacies.yaml +5 -0
- lm-evaluation-harness/lm_eval/tasks/global_mmlu/full/mg/global_mmlu_full_mg_machine_learning.yaml +5 -0
- lm-evaluation-harness/lm_eval/tasks/global_mmlu/full/mg/global_mmlu_full_mg_management.yaml +5 -0
- lm-evaluation-harness/lm_eval/tasks/global_mmlu/full/mg/global_mmlu_full_mg_marketing.yaml +5 -0
- lm-evaluation-harness/lm_eval/tasks/global_mmlu/full/mg/global_mmlu_full_mg_nutrition.yaml +5 -0
- lm-evaluation-harness/lm_eval/tasks/global_mmlu/full/mg/global_mmlu_full_mg_philosophy.yaml +5 -0
- lm-evaluation-harness/lm_eval/tasks/global_mmlu/full/mg/global_mmlu_full_mg_prehistory.yaml +5 -0
- lm-evaluation-harness/lm_eval/tasks/global_mmlu/full/mg/global_mmlu_full_mg_professional_accounting.yaml +5 -0
- lm-evaluation-harness/lm_eval/tasks/global_mmlu/full/mg/global_mmlu_full_mg_professional_law.yaml +5 -0
- lm-evaluation-harness/lm_eval/tasks/global_mmlu/full/mg/global_mmlu_full_mg_professional_psychology.yaml +5 -0
- lm-evaluation-harness/lm_eval/tasks/global_mmlu/full/mg/global_mmlu_full_mg_public_relations.yaml +5 -0
- lm-evaluation-harness/lm_eval/tasks/global_mmlu/full/mg/global_mmlu_full_mg_security_studies.yaml +5 -0
- lm-evaluation-harness/lm_eval/tasks/global_mmlu/full/mg/global_mmlu_full_mg_sociology.yaml +5 -0
- lm-evaluation-harness/lm_eval/tasks/global_mmlu/full/mg/global_mmlu_full_mg_us_foreign_policy.yaml +5 -0
- lm-evaluation-harness/lm_eval/tasks/global_mmlu/full/mg/global_mmlu_full_mg_virology.yaml +5 -0
- lm-evaluation-harness/lm_eval/tasks/global_mmlu/full/mg/global_mmlu_full_mg_world_religions.yaml +5 -0
- lm-evaluation-harness/lm_eval/tasks/global_mmlu/full/mg/utils.py +73 -0
- lm-evaluation-harness/lm_eval/tasks/global_mmlu/full/ms/_global_mmlu_full_ms.yaml +11 -0
- lm-evaluation-harness/lm_eval/tasks/global_mmlu/full/ms/_global_mmlu_full_ms_humanities.yaml +8 -0
- lm-evaluation-harness/lm_eval/tasks/global_mmlu/full/ms/_global_mmlu_full_ms_other.yaml +8 -0
- lm-evaluation-harness/lm_eval/tasks/global_mmlu/full/ms/_global_mmlu_full_ms_stem.yaml +8 -0
- lm-evaluation-harness/lm_eval/tasks/global_mmlu/full/ms/_ms_template_yaml +16 -0
- lm-evaluation-harness/lm_eval/tasks/global_mmlu/full/ms/global_mmlu_full_ms_abstract_algebra.yaml +5 -0
- lm-evaluation-harness/lm_eval/tasks/global_mmlu/full/ms/global_mmlu_full_ms_astronomy.yaml +5 -0
- lm-evaluation-harness/lm_eval/tasks/global_mmlu/full/ms/global_mmlu_full_ms_clinical_knowledge.yaml +5 -0
- lm-evaluation-harness/lm_eval/tasks/global_mmlu/full/ms/global_mmlu_full_ms_college_chemistry.yaml +5 -0
- lm-evaluation-harness/lm_eval/tasks/global_mmlu/full/ms/global_mmlu_full_ms_college_computer_science.yaml +5 -0
- lm-evaluation-harness/lm_eval/tasks/global_mmlu/full/ms/global_mmlu_full_ms_college_mathematics.yaml +5 -0
- lm-evaluation-harness/lm_eval/tasks/global_mmlu/full/ms/global_mmlu_full_ms_college_medicine.yaml +5 -0
- lm-evaluation-harness/lm_eval/tasks/global_mmlu/full/ms/global_mmlu_full_ms_college_physics.yaml +5 -0
- lm-evaluation-harness/lm_eval/tasks/global_mmlu/full/ms/global_mmlu_full_ms_computer_security.yaml +5 -0
- lm-evaluation-harness/lm_eval/tasks/global_mmlu/full/ms/global_mmlu_full_ms_conceptual_physics.yaml +5 -0
- lm-evaluation-harness/lm_eval/tasks/global_mmlu/full/ms/global_mmlu_full_ms_econometrics.yaml +5 -0
- lm-evaluation-harness/lm_eval/tasks/global_mmlu/full/ms/global_mmlu_full_ms_electrical_engineering.yaml +5 -0
- lm-evaluation-harness/lm_eval/tasks/global_mmlu/full/ms/global_mmlu_full_ms_elementary_mathematics.yaml +5 -0
- lm-evaluation-harness/lm_eval/tasks/global_mmlu/full/ms/global_mmlu_full_ms_formal_logic.yaml +5 -0
- lm-evaluation-harness/lm_eval/tasks/global_mmlu/full/ms/global_mmlu_full_ms_global_facts.yaml +5 -0
- lm-evaluation-harness/lm_eval/tasks/global_mmlu/full/ms/global_mmlu_full_ms_high_school_biology.yaml +5 -0
- lm-evaluation-harness/lm_eval/tasks/global_mmlu/full/ms/global_mmlu_full_ms_high_school_chemistry.yaml +5 -0
- lm-evaluation-harness/lm_eval/tasks/global_mmlu/full/ms/global_mmlu_full_ms_high_school_computer_science.yaml +5 -0
- lm-evaluation-harness/lm_eval/tasks/global_mmlu/full/ms/global_mmlu_full_ms_high_school_european_history.yaml +5 -0
- lm-evaluation-harness/lm_eval/tasks/global_mmlu/full/ms/global_mmlu_full_ms_high_school_geography.yaml +5 -0
- lm-evaluation-harness/lm_eval/tasks/global_mmlu/full/ms/global_mmlu_full_ms_high_school_government_and_politics.yaml +5 -0
- lm-evaluation-harness/lm_eval/tasks/global_mmlu/full/ms/global_mmlu_full_ms_high_school_macroeconomics.yaml +5 -0
- lm-evaluation-harness/lm_eval/tasks/global_mmlu/full/ms/global_mmlu_full_ms_high_school_mathematics.yaml +5 -0
- lm-evaluation-harness/lm_eval/tasks/global_mmlu/full/ms/global_mmlu_full_ms_high_school_microeconomics.yaml +5 -0
- lm-evaluation-harness/lm_eval/tasks/global_mmlu/full/ms/global_mmlu_full_ms_high_school_physics.yaml +5 -0
- lm-evaluation-harness/lm_eval/tasks/global_mmlu/full/ms/global_mmlu_full_ms_high_school_psychology.yaml +5 -0
lm-evaluation-harness/lm_eval/tasks/global_mmlu/full/mg/global_mmlu_full_mg_human_sexuality.yaml
ADDED
|
@@ -0,0 +1,5 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# Generated by _generate_configs.py
|
| 2 |
+
include: _mg_template_yaml
|
| 3 |
+
process_docs: !function utils.process_human_sexuality
|
| 4 |
+
tag: global_mmlu_full_mg_social_sciences_tasks
|
| 5 |
+
task: global_mmlu_full_mg_human_sexuality
|
lm-evaluation-harness/lm_eval/tasks/global_mmlu/full/mg/global_mmlu_full_mg_jurisprudence.yaml
ADDED
|
@@ -0,0 +1,5 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# Generated by _generate_configs.py
|
| 2 |
+
include: _mg_template_yaml
|
| 3 |
+
process_docs: !function utils.process_jurisprudence
|
| 4 |
+
tag: global_mmlu_full_mg_humanities_tasks
|
| 5 |
+
task: global_mmlu_full_mg_jurisprudence
|
lm-evaluation-harness/lm_eval/tasks/global_mmlu/full/mg/global_mmlu_full_mg_logical_fallacies.yaml
ADDED
|
@@ -0,0 +1,5 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# Generated by _generate_configs.py
|
| 2 |
+
include: _mg_template_yaml
|
| 3 |
+
process_docs: !function utils.process_logical_fallacies
|
| 4 |
+
tag: global_mmlu_full_mg_humanities_tasks
|
| 5 |
+
task: global_mmlu_full_mg_logical_fallacies
|
lm-evaluation-harness/lm_eval/tasks/global_mmlu/full/mg/global_mmlu_full_mg_machine_learning.yaml
ADDED
|
@@ -0,0 +1,5 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# Generated by _generate_configs.py
|
| 2 |
+
include: _mg_template_yaml
|
| 3 |
+
process_docs: !function utils.process_machine_learning
|
| 4 |
+
tag: global_mmlu_full_mg_stem_tasks
|
| 5 |
+
task: global_mmlu_full_mg_machine_learning
|
lm-evaluation-harness/lm_eval/tasks/global_mmlu/full/mg/global_mmlu_full_mg_management.yaml
ADDED
|
@@ -0,0 +1,5 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# Generated by _generate_configs.py
|
| 2 |
+
include: _mg_template_yaml
|
| 3 |
+
process_docs: !function utils.process_management
|
| 4 |
+
tag: global_mmlu_full_mg_other_tasks
|
| 5 |
+
task: global_mmlu_full_mg_management
|
lm-evaluation-harness/lm_eval/tasks/global_mmlu/full/mg/global_mmlu_full_mg_marketing.yaml
ADDED
|
@@ -0,0 +1,5 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# Generated by _generate_configs.py
|
| 2 |
+
include: _mg_template_yaml
|
| 3 |
+
process_docs: !function utils.process_marketing
|
| 4 |
+
tag: global_mmlu_full_mg_other_tasks
|
| 5 |
+
task: global_mmlu_full_mg_marketing
|
lm-evaluation-harness/lm_eval/tasks/global_mmlu/full/mg/global_mmlu_full_mg_nutrition.yaml
ADDED
|
@@ -0,0 +1,5 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# Generated by _generate_configs.py
|
| 2 |
+
include: _mg_template_yaml
|
| 3 |
+
process_docs: !function utils.process_nutrition
|
| 4 |
+
tag: global_mmlu_full_mg_other_tasks
|
| 5 |
+
task: global_mmlu_full_mg_nutrition
|
lm-evaluation-harness/lm_eval/tasks/global_mmlu/full/mg/global_mmlu_full_mg_philosophy.yaml
ADDED
|
@@ -0,0 +1,5 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# Generated by _generate_configs.py
|
| 2 |
+
include: _mg_template_yaml
|
| 3 |
+
process_docs: !function utils.process_philosophy
|
| 4 |
+
tag: global_mmlu_full_mg_humanities_tasks
|
| 5 |
+
task: global_mmlu_full_mg_philosophy
|
lm-evaluation-harness/lm_eval/tasks/global_mmlu/full/mg/global_mmlu_full_mg_prehistory.yaml
ADDED
|
@@ -0,0 +1,5 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# Generated by _generate_configs.py
|
| 2 |
+
include: _mg_template_yaml
|
| 3 |
+
process_docs: !function utils.process_prehistory
|
| 4 |
+
tag: global_mmlu_full_mg_humanities_tasks
|
| 5 |
+
task: global_mmlu_full_mg_prehistory
|
lm-evaluation-harness/lm_eval/tasks/global_mmlu/full/mg/global_mmlu_full_mg_professional_accounting.yaml
ADDED
|
@@ -0,0 +1,5 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# Generated by _generate_configs.py
|
| 2 |
+
include: _mg_template_yaml
|
| 3 |
+
process_docs: !function utils.process_professional_accounting
|
| 4 |
+
tag: global_mmlu_full_mg_other_tasks
|
| 5 |
+
task: global_mmlu_full_mg_professional_accounting
|
lm-evaluation-harness/lm_eval/tasks/global_mmlu/full/mg/global_mmlu_full_mg_professional_law.yaml
ADDED
|
@@ -0,0 +1,5 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# Generated by _generate_configs.py
|
| 2 |
+
include: _mg_template_yaml
|
| 3 |
+
process_docs: !function utils.process_professional_law
|
| 4 |
+
tag: global_mmlu_full_mg_humanities_tasks
|
| 5 |
+
task: global_mmlu_full_mg_professional_law
|
lm-evaluation-harness/lm_eval/tasks/global_mmlu/full/mg/global_mmlu_full_mg_professional_psychology.yaml
ADDED
|
@@ -0,0 +1,5 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# Generated by _generate_configs.py
|
| 2 |
+
include: _mg_template_yaml
|
| 3 |
+
process_docs: !function utils.process_professional_psychology
|
| 4 |
+
tag: global_mmlu_full_mg_social_sciences_tasks
|
| 5 |
+
task: global_mmlu_full_mg_professional_psychology
|
lm-evaluation-harness/lm_eval/tasks/global_mmlu/full/mg/global_mmlu_full_mg_public_relations.yaml
ADDED
|
@@ -0,0 +1,5 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# Generated by _generate_configs.py
|
| 2 |
+
include: _mg_template_yaml
|
| 3 |
+
process_docs: !function utils.process_public_relations
|
| 4 |
+
tag: global_mmlu_full_mg_social_sciences_tasks
|
| 5 |
+
task: global_mmlu_full_mg_public_relations
|
lm-evaluation-harness/lm_eval/tasks/global_mmlu/full/mg/global_mmlu_full_mg_security_studies.yaml
ADDED
|
@@ -0,0 +1,5 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# Generated by _generate_configs.py
|
| 2 |
+
include: _mg_template_yaml
|
| 3 |
+
process_docs: !function utils.process_security_studies
|
| 4 |
+
tag: global_mmlu_full_mg_social_sciences_tasks
|
| 5 |
+
task: global_mmlu_full_mg_security_studies
|
lm-evaluation-harness/lm_eval/tasks/global_mmlu/full/mg/global_mmlu_full_mg_sociology.yaml
ADDED
|
@@ -0,0 +1,5 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# Generated by _generate_configs.py
|
| 2 |
+
include: _mg_template_yaml
|
| 3 |
+
process_docs: !function utils.process_sociology
|
| 4 |
+
tag: global_mmlu_full_mg_social_sciences_tasks
|
| 5 |
+
task: global_mmlu_full_mg_sociology
|
lm-evaluation-harness/lm_eval/tasks/global_mmlu/full/mg/global_mmlu_full_mg_us_foreign_policy.yaml
ADDED
|
@@ -0,0 +1,5 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# Generated by _generate_configs.py
|
| 2 |
+
include: _mg_template_yaml
|
| 3 |
+
process_docs: !function utils.process_us_foreign_policy
|
| 4 |
+
tag: global_mmlu_full_mg_social_sciences_tasks
|
| 5 |
+
task: global_mmlu_full_mg_us_foreign_policy
|
lm-evaluation-harness/lm_eval/tasks/global_mmlu/full/mg/global_mmlu_full_mg_virology.yaml
ADDED
|
@@ -0,0 +1,5 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# Generated by _generate_configs.py
|
| 2 |
+
include: _mg_template_yaml
|
| 3 |
+
process_docs: !function utils.process_virology
|
| 4 |
+
tag: global_mmlu_full_mg_other_tasks
|
| 5 |
+
task: global_mmlu_full_mg_virology
|
lm-evaluation-harness/lm_eval/tasks/global_mmlu/full/mg/global_mmlu_full_mg_world_religions.yaml
ADDED
|
@@ -0,0 +1,5 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# Generated by _generate_configs.py
|
| 2 |
+
include: _mg_template_yaml
|
| 3 |
+
process_docs: !function utils.process_world_religions
|
| 4 |
+
tag: global_mmlu_full_mg_humanities_tasks
|
| 5 |
+
task: global_mmlu_full_mg_world_religions
|
lm-evaluation-harness/lm_eval/tasks/global_mmlu/full/mg/utils.py
ADDED
|
@@ -0,0 +1,73 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
from functools import partial
|
| 2 |
+
|
| 3 |
+
|
| 4 |
+
SUBJECTS = [
|
| 5 |
+
"abstract_algebra",
|
| 6 |
+
"anatomy",
|
| 7 |
+
"astronomy",
|
| 8 |
+
"business_ethics",
|
| 9 |
+
"clinical_knowledge",
|
| 10 |
+
"college_biology",
|
| 11 |
+
"college_chemistry",
|
| 12 |
+
"college_computer_science",
|
| 13 |
+
"college_mathematics",
|
| 14 |
+
"college_medicine",
|
| 15 |
+
"college_physics",
|
| 16 |
+
"computer_security",
|
| 17 |
+
"conceptual_physics",
|
| 18 |
+
"econometrics",
|
| 19 |
+
"electrical_engineering",
|
| 20 |
+
"elementary_mathematics",
|
| 21 |
+
"formal_logic",
|
| 22 |
+
"global_facts",
|
| 23 |
+
"high_school_biology",
|
| 24 |
+
"high_school_chemistry",
|
| 25 |
+
"high_school_computer_science",
|
| 26 |
+
"high_school_european_history",
|
| 27 |
+
"high_school_geography",
|
| 28 |
+
"high_school_government_and_politics",
|
| 29 |
+
"high_school_macroeconomics",
|
| 30 |
+
"high_school_mathematics",
|
| 31 |
+
"high_school_microeconomics",
|
| 32 |
+
"high_school_physics",
|
| 33 |
+
"high_school_psychology",
|
| 34 |
+
"high_school_statistics",
|
| 35 |
+
"high_school_us_history",
|
| 36 |
+
"high_school_world_history",
|
| 37 |
+
"human_aging",
|
| 38 |
+
"human_sexuality",
|
| 39 |
+
"international_law",
|
| 40 |
+
"jurisprudence",
|
| 41 |
+
"logical_fallacies",
|
| 42 |
+
"machine_learning",
|
| 43 |
+
"management",
|
| 44 |
+
"marketing",
|
| 45 |
+
"medical_genetics",
|
| 46 |
+
"miscellaneous",
|
| 47 |
+
"moral_disputes",
|
| 48 |
+
"moral_scenarios",
|
| 49 |
+
"nutrition",
|
| 50 |
+
"philosophy",
|
| 51 |
+
"prehistory",
|
| 52 |
+
"professional_accounting",
|
| 53 |
+
"professional_law",
|
| 54 |
+
"professional_medicine",
|
| 55 |
+
"professional_psychology",
|
| 56 |
+
"public_relations",
|
| 57 |
+
"security_studies",
|
| 58 |
+
"sociology",
|
| 59 |
+
"us_foreign_policy",
|
| 60 |
+
"virology",
|
| 61 |
+
"world_religions",
|
| 62 |
+
]
|
| 63 |
+
|
| 64 |
+
|
| 65 |
+
def process_docs(dataset, subject):
|
| 66 |
+
return dataset.filter(lambda x: x["subject"] == subject)
|
| 67 |
+
|
| 68 |
+
|
| 69 |
+
process_functions = {
|
| 70 |
+
f"process_{subject}": partial(process_docs, subject=subject) for subject in SUBJECTS
|
| 71 |
+
}
|
| 72 |
+
|
| 73 |
+
globals().update(process_functions)
|
lm-evaluation-harness/lm_eval/tasks/global_mmlu/full/ms/_global_mmlu_full_ms.yaml
ADDED
|
@@ -0,0 +1,11 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
group: global_mmlu_full_ms
|
| 2 |
+
task:
|
| 3 |
+
- global_mmlu_full_ms_stem
|
| 4 |
+
- global_mmlu_full_ms_other
|
| 5 |
+
- global_mmlu_full_ms_social_sciences
|
| 6 |
+
- global_mmlu_full_ms_humanities
|
| 7 |
+
aggregate_metric_list:
|
| 8 |
+
- metric: acc
|
| 9 |
+
weight_by_size: True
|
| 10 |
+
metadata:
|
| 11 |
+
version: 0.0
|
lm-evaluation-harness/lm_eval/tasks/global_mmlu/full/ms/_global_mmlu_full_ms_humanities.yaml
ADDED
|
@@ -0,0 +1,8 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
group: global_mmlu_full_ms_humanities
|
| 2 |
+
task:
|
| 3 |
+
- global_mmlu_full_ms_humanities_tasks
|
| 4 |
+
aggregate_metric_list:
|
| 5 |
+
- metric: acc
|
| 6 |
+
weight_by_size: true
|
| 7 |
+
metadata:
|
| 8 |
+
version: 0.0
|
lm-evaluation-harness/lm_eval/tasks/global_mmlu/full/ms/_global_mmlu_full_ms_other.yaml
ADDED
|
@@ -0,0 +1,8 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
group: global_mmlu_full_ms_other
|
| 2 |
+
task:
|
| 3 |
+
- global_mmlu_full_ms_other_tasks
|
| 4 |
+
aggregate_metric_list:
|
| 5 |
+
- metric: acc
|
| 6 |
+
weight_by_size: true
|
| 7 |
+
metadata:
|
| 8 |
+
version: 0.0
|
lm-evaluation-harness/lm_eval/tasks/global_mmlu/full/ms/_global_mmlu_full_ms_stem.yaml
ADDED
|
@@ -0,0 +1,8 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
group: global_mmlu_full_ms_stem
|
| 2 |
+
task:
|
| 3 |
+
- global_mmlu_full_ms_stem_tasks
|
| 4 |
+
aggregate_metric_list:
|
| 5 |
+
- metric: acc
|
| 6 |
+
weight_by_size: true
|
| 7 |
+
metadata:
|
| 8 |
+
version: 0.0
|
lm-evaluation-harness/lm_eval/tasks/global_mmlu/full/ms/_ms_template_yaml
ADDED
|
@@ -0,0 +1,16 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
dataset_path: CohereForAI/Global-MMLU
|
| 2 |
+
dataset_name: ms
|
| 3 |
+
test_split: test
|
| 4 |
+
fewshot_split: dev
|
| 5 |
+
fewshot_config:
|
| 6 |
+
sampler: first_n
|
| 7 |
+
output_type: multiple_choice
|
| 8 |
+
doc_to_text: "{{question.strip()}}\nA. {{option_a}}\nB. {{option_b}}\nC. {{option_c}}\nD. {{option_d}}\nAnswer:"
|
| 9 |
+
doc_to_choice: ["A", "B", "C", "D"]
|
| 10 |
+
doc_to_target: answer
|
| 11 |
+
metric_list:
|
| 12 |
+
- metric: acc
|
| 13 |
+
aggregation: mean
|
| 14 |
+
higher_is_better: true
|
| 15 |
+
metadata:
|
| 16 |
+
version: 0.0
|
lm-evaluation-harness/lm_eval/tasks/global_mmlu/full/ms/global_mmlu_full_ms_abstract_algebra.yaml
ADDED
|
@@ -0,0 +1,5 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# Generated by _generate_configs.py
|
| 2 |
+
include: _ms_template_yaml
|
| 3 |
+
process_docs: !function utils.process_abstract_algebra
|
| 4 |
+
tag: global_mmlu_full_ms_stem_tasks
|
| 5 |
+
task: global_mmlu_full_ms_abstract_algebra
|
lm-evaluation-harness/lm_eval/tasks/global_mmlu/full/ms/global_mmlu_full_ms_astronomy.yaml
ADDED
|
@@ -0,0 +1,5 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# Generated by _generate_configs.py
|
| 2 |
+
include: _ms_template_yaml
|
| 3 |
+
process_docs: !function utils.process_astronomy
|
| 4 |
+
tag: global_mmlu_full_ms_stem_tasks
|
| 5 |
+
task: global_mmlu_full_ms_astronomy
|
lm-evaluation-harness/lm_eval/tasks/global_mmlu/full/ms/global_mmlu_full_ms_clinical_knowledge.yaml
ADDED
|
@@ -0,0 +1,5 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# Generated by _generate_configs.py
|
| 2 |
+
include: _ms_template_yaml
|
| 3 |
+
process_docs: !function utils.process_clinical_knowledge
|
| 4 |
+
tag: global_mmlu_full_ms_other_tasks
|
| 5 |
+
task: global_mmlu_full_ms_clinical_knowledge
|
lm-evaluation-harness/lm_eval/tasks/global_mmlu/full/ms/global_mmlu_full_ms_college_chemistry.yaml
ADDED
|
@@ -0,0 +1,5 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# Generated by _generate_configs.py
|
| 2 |
+
include: _ms_template_yaml
|
| 3 |
+
process_docs: !function utils.process_college_chemistry
|
| 4 |
+
tag: global_mmlu_full_ms_stem_tasks
|
| 5 |
+
task: global_mmlu_full_ms_college_chemistry
|
lm-evaluation-harness/lm_eval/tasks/global_mmlu/full/ms/global_mmlu_full_ms_college_computer_science.yaml
ADDED
|
@@ -0,0 +1,5 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# Generated by _generate_configs.py
|
| 2 |
+
include: _ms_template_yaml
|
| 3 |
+
process_docs: !function utils.process_college_computer_science
|
| 4 |
+
tag: global_mmlu_full_ms_stem_tasks
|
| 5 |
+
task: global_mmlu_full_ms_college_computer_science
|
lm-evaluation-harness/lm_eval/tasks/global_mmlu/full/ms/global_mmlu_full_ms_college_mathematics.yaml
ADDED
|
@@ -0,0 +1,5 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# Generated by _generate_configs.py
|
| 2 |
+
include: _ms_template_yaml
|
| 3 |
+
process_docs: !function utils.process_college_mathematics
|
| 4 |
+
tag: global_mmlu_full_ms_stem_tasks
|
| 5 |
+
task: global_mmlu_full_ms_college_mathematics
|
lm-evaluation-harness/lm_eval/tasks/global_mmlu/full/ms/global_mmlu_full_ms_college_medicine.yaml
ADDED
|
@@ -0,0 +1,5 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# Generated by _generate_configs.py
|
| 2 |
+
include: _ms_template_yaml
|
| 3 |
+
process_docs: !function utils.process_college_medicine
|
| 4 |
+
tag: global_mmlu_full_ms_other_tasks
|
| 5 |
+
task: global_mmlu_full_ms_college_medicine
|
lm-evaluation-harness/lm_eval/tasks/global_mmlu/full/ms/global_mmlu_full_ms_college_physics.yaml
ADDED
|
@@ -0,0 +1,5 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# Generated by _generate_configs.py
|
| 2 |
+
include: _ms_template_yaml
|
| 3 |
+
process_docs: !function utils.process_college_physics
|
| 4 |
+
tag: global_mmlu_full_ms_stem_tasks
|
| 5 |
+
task: global_mmlu_full_ms_college_physics
|
lm-evaluation-harness/lm_eval/tasks/global_mmlu/full/ms/global_mmlu_full_ms_computer_security.yaml
ADDED
|
@@ -0,0 +1,5 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# Generated by _generate_configs.py
|
| 2 |
+
include: _ms_template_yaml
|
| 3 |
+
process_docs: !function utils.process_computer_security
|
| 4 |
+
tag: global_mmlu_full_ms_stem_tasks
|
| 5 |
+
task: global_mmlu_full_ms_computer_security
|
lm-evaluation-harness/lm_eval/tasks/global_mmlu/full/ms/global_mmlu_full_ms_conceptual_physics.yaml
ADDED
|
@@ -0,0 +1,5 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# Generated by _generate_configs.py
|
| 2 |
+
include: _ms_template_yaml
|
| 3 |
+
process_docs: !function utils.process_conceptual_physics
|
| 4 |
+
tag: global_mmlu_full_ms_stem_tasks
|
| 5 |
+
task: global_mmlu_full_ms_conceptual_physics
|
lm-evaluation-harness/lm_eval/tasks/global_mmlu/full/ms/global_mmlu_full_ms_econometrics.yaml
ADDED
|
@@ -0,0 +1,5 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# Generated by _generate_configs.py
|
| 2 |
+
include: _ms_template_yaml
|
| 3 |
+
process_docs: !function utils.process_econometrics
|
| 4 |
+
tag: global_mmlu_full_ms_social_sciences_tasks
|
| 5 |
+
task: global_mmlu_full_ms_econometrics
|
lm-evaluation-harness/lm_eval/tasks/global_mmlu/full/ms/global_mmlu_full_ms_electrical_engineering.yaml
ADDED
|
@@ -0,0 +1,5 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# Generated by _generate_configs.py
|
| 2 |
+
include: _ms_template_yaml
|
| 3 |
+
process_docs: !function utils.process_electrical_engineering
|
| 4 |
+
tag: global_mmlu_full_ms_stem_tasks
|
| 5 |
+
task: global_mmlu_full_ms_electrical_engineering
|
lm-evaluation-harness/lm_eval/tasks/global_mmlu/full/ms/global_mmlu_full_ms_elementary_mathematics.yaml
ADDED
|
@@ -0,0 +1,5 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# Generated by _generate_configs.py
|
| 2 |
+
include: _ms_template_yaml
|
| 3 |
+
process_docs: !function utils.process_elementary_mathematics
|
| 4 |
+
tag: global_mmlu_full_ms_stem_tasks
|
| 5 |
+
task: global_mmlu_full_ms_elementary_mathematics
|
lm-evaluation-harness/lm_eval/tasks/global_mmlu/full/ms/global_mmlu_full_ms_formal_logic.yaml
ADDED
|
@@ -0,0 +1,5 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# Generated by _generate_configs.py
|
| 2 |
+
include: _ms_template_yaml
|
| 3 |
+
process_docs: !function utils.process_formal_logic
|
| 4 |
+
tag: global_mmlu_full_ms_humanities_tasks
|
| 5 |
+
task: global_mmlu_full_ms_formal_logic
|
lm-evaluation-harness/lm_eval/tasks/global_mmlu/full/ms/global_mmlu_full_ms_global_facts.yaml
ADDED
|
@@ -0,0 +1,5 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# Generated by _generate_configs.py
|
| 2 |
+
include: _ms_template_yaml
|
| 3 |
+
process_docs: !function utils.process_global_facts
|
| 4 |
+
tag: global_mmlu_full_ms_other_tasks
|
| 5 |
+
task: global_mmlu_full_ms_global_facts
|
lm-evaluation-harness/lm_eval/tasks/global_mmlu/full/ms/global_mmlu_full_ms_high_school_biology.yaml
ADDED
|
@@ -0,0 +1,5 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# Generated by _generate_configs.py
|
| 2 |
+
include: _ms_template_yaml
|
| 3 |
+
process_docs: !function utils.process_high_school_biology
|
| 4 |
+
tag: global_mmlu_full_ms_stem_tasks
|
| 5 |
+
task: global_mmlu_full_ms_high_school_biology
|
lm-evaluation-harness/lm_eval/tasks/global_mmlu/full/ms/global_mmlu_full_ms_high_school_chemistry.yaml
ADDED
|
@@ -0,0 +1,5 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# Generated by _generate_configs.py
|
| 2 |
+
include: _ms_template_yaml
|
| 3 |
+
process_docs: !function utils.process_high_school_chemistry
|
| 4 |
+
tag: global_mmlu_full_ms_stem_tasks
|
| 5 |
+
task: global_mmlu_full_ms_high_school_chemistry
|
lm-evaluation-harness/lm_eval/tasks/global_mmlu/full/ms/global_mmlu_full_ms_high_school_computer_science.yaml
ADDED
|
@@ -0,0 +1,5 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# Generated by _generate_configs.py
|
| 2 |
+
include: _ms_template_yaml
|
| 3 |
+
process_docs: !function utils.process_high_school_computer_science
|
| 4 |
+
tag: global_mmlu_full_ms_stem_tasks
|
| 5 |
+
task: global_mmlu_full_ms_high_school_computer_science
|
lm-evaluation-harness/lm_eval/tasks/global_mmlu/full/ms/global_mmlu_full_ms_high_school_european_history.yaml
ADDED
|
@@ -0,0 +1,5 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# Generated by _generate_configs.py
|
| 2 |
+
include: _ms_template_yaml
|
| 3 |
+
process_docs: !function utils.process_high_school_european_history
|
| 4 |
+
tag: global_mmlu_full_ms_humanities_tasks
|
| 5 |
+
task: global_mmlu_full_ms_high_school_european_history
|
lm-evaluation-harness/lm_eval/tasks/global_mmlu/full/ms/global_mmlu_full_ms_high_school_geography.yaml
ADDED
|
@@ -0,0 +1,5 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# Generated by _generate_configs.py
|
| 2 |
+
include: _ms_template_yaml
|
| 3 |
+
process_docs: !function utils.process_high_school_geography
|
| 4 |
+
tag: global_mmlu_full_ms_social_sciences_tasks
|
| 5 |
+
task: global_mmlu_full_ms_high_school_geography
|
lm-evaluation-harness/lm_eval/tasks/global_mmlu/full/ms/global_mmlu_full_ms_high_school_government_and_politics.yaml
ADDED
|
@@ -0,0 +1,5 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# Generated by _generate_configs.py
|
| 2 |
+
include: _ms_template_yaml
|
| 3 |
+
process_docs: !function utils.process_high_school_government_and_politics
|
| 4 |
+
tag: global_mmlu_full_ms_social_sciences_tasks
|
| 5 |
+
task: global_mmlu_full_ms_high_school_government_and_politics
|
lm-evaluation-harness/lm_eval/tasks/global_mmlu/full/ms/global_mmlu_full_ms_high_school_macroeconomics.yaml
ADDED
|
@@ -0,0 +1,5 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# Generated by _generate_configs.py
|
| 2 |
+
include: _ms_template_yaml
|
| 3 |
+
process_docs: !function utils.process_high_school_macroeconomics
|
| 4 |
+
tag: global_mmlu_full_ms_social_sciences_tasks
|
| 5 |
+
task: global_mmlu_full_ms_high_school_macroeconomics
|
lm-evaluation-harness/lm_eval/tasks/global_mmlu/full/ms/global_mmlu_full_ms_high_school_mathematics.yaml
ADDED
|
@@ -0,0 +1,5 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# Generated by _generate_configs.py
|
| 2 |
+
include: _ms_template_yaml
|
| 3 |
+
process_docs: !function utils.process_high_school_mathematics
|
| 4 |
+
tag: global_mmlu_full_ms_stem_tasks
|
| 5 |
+
task: global_mmlu_full_ms_high_school_mathematics
|
lm-evaluation-harness/lm_eval/tasks/global_mmlu/full/ms/global_mmlu_full_ms_high_school_microeconomics.yaml
ADDED
|
@@ -0,0 +1,5 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# Generated by _generate_configs.py
|
| 2 |
+
include: _ms_template_yaml
|
| 3 |
+
process_docs: !function utils.process_high_school_microeconomics
|
| 4 |
+
tag: global_mmlu_full_ms_social_sciences_tasks
|
| 5 |
+
task: global_mmlu_full_ms_high_school_microeconomics
|
lm-evaluation-harness/lm_eval/tasks/global_mmlu/full/ms/global_mmlu_full_ms_high_school_physics.yaml
ADDED
|
@@ -0,0 +1,5 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# Generated by _generate_configs.py
|
| 2 |
+
include: _ms_template_yaml
|
| 3 |
+
process_docs: !function utils.process_high_school_physics
|
| 4 |
+
tag: global_mmlu_full_ms_stem_tasks
|
| 5 |
+
task: global_mmlu_full_ms_high_school_physics
|
lm-evaluation-harness/lm_eval/tasks/global_mmlu/full/ms/global_mmlu_full_ms_high_school_psychology.yaml
ADDED
|
@@ -0,0 +1,5 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# Generated by _generate_configs.py
|
| 2 |
+
include: _ms_template_yaml
|
| 3 |
+
process_docs: !function utils.process_high_school_psychology
|
| 4 |
+
tag: global_mmlu_full_ms_social_sciences_tasks
|
| 5 |
+
task: global_mmlu_full_ms_high_school_psychology
|