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/include/few_shot_en/Serbian/utils.py +30 -0
- lm-evaluation-harness/lm_eval/tasks/include/few_shot_en/Spanish/include_base_44_spanish_health_oriented_education.yaml +5 -0
- lm-evaluation-harness/lm_eval/tasks/include/few_shot_en/Spanish/include_base_44_spanish_stem.yaml +4 -0
- lm-evaluation-harness/lm_eval/tasks/include/few_shot_en/Spanish/utils.py +30 -0
- lm-evaluation-harness/lm_eval/tasks/include/few_shot_en/Tagalog/_tagalog_few_shot_en_template_yaml +18 -0
- lm-evaluation-harness/lm_eval/tasks/include/few_shot_en/Tagalog/include_base_44_tagalog_arts_humanities.yaml +5 -0
- lm-evaluation-harness/lm_eval/tasks/include/few_shot_en/Tamil/_include_base_44_tamil.yaml +9 -0
- lm-evaluation-harness/lm_eval/tasks/include/few_shot_en/Tamil/include_base_44_tamil_general_knowledge.yaml +5 -0
- lm-evaluation-harness/lm_eval/tasks/include/few_shot_en/Tamil/include_base_44_tamil_stem.yaml +4 -0
- lm-evaluation-harness/lm_eval/tasks/include/few_shot_en/Telugu/_include_base_44_telugu.yaml +11 -0
- lm-evaluation-harness/lm_eval/tasks/include/few_shot_en/Telugu/include_base_44_telugu_arts_humanities.yaml +5 -0
- lm-evaluation-harness/lm_eval/tasks/include/few_shot_en/Telugu/include_base_44_telugu_social_science.yaml +5 -0
- lm-evaluation-harness/lm_eval/tasks/include/few_shot_en/Telugu/utils.py +30 -0
- lm-evaluation-harness/lm_eval/tasks/include/few_shot_en/Turkish/_include_base_44_turkish.yaml +11 -0
- lm-evaluation-harness/lm_eval/tasks/include/few_shot_en/Turkish/include_base_44_turkish_business_commerce.yaml +5 -0
- lm-evaluation-harness/lm_eval/tasks/include/few_shot_en/Turkish/include_base_44_turkish_social_science.yaml +5 -0
- lm-evaluation-harness/lm_eval/tasks/include/few_shot_en/Turkish/include_base_44_turkish_stem.yaml +4 -0
- lm-evaluation-harness/lm_eval/tasks/include/few_shot_en/Turkish/utils.py +30 -0
- lm-evaluation-harness/lm_eval/tasks/include/few_shot_en/Ukrainian/_include_base_44_ukrainian.yaml +10 -0
- lm-evaluation-harness/lm_eval/tasks/include/few_shot_en/Ukrainian/_ukrainian_few_shot_en_template_yaml +18 -0
- lm-evaluation-harness/lm_eval/tasks/include/few_shot_en/Ukrainian/include_base_44_ukrainian_arts_humanities.yaml +5 -0
- lm-evaluation-harness/lm_eval/tasks/include/few_shot_en/Ukrainian/include_base_44_ukrainian_social_science.yaml +5 -0
- lm-evaluation-harness/lm_eval/tasks/include/few_shot_en/Ukrainian/include_base_44_ukrainian_stem.yaml +4 -0
- lm-evaluation-harness/lm_eval/tasks/include/few_shot_en/Ukrainian/utils.py +30 -0
- lm-evaluation-harness/lm_eval/tasks/include/few_shot_en/Urdu/_include_base_44_urdu.yaml +10 -0
- lm-evaluation-harness/lm_eval/tasks/include/few_shot_en/Urdu/_urdu_few_shot_en_template_yaml +18 -0
- lm-evaluation-harness/lm_eval/tasks/include/few_shot_en/Urdu/include_base_44_urdu_arts_humanities.yaml +5 -0
- lm-evaluation-harness/lm_eval/tasks/include/few_shot_en/Urdu/include_base_44_urdu_health_oriented_education.yaml +5 -0
- lm-evaluation-harness/lm_eval/tasks/include/few_shot_en/Urdu/include_base_44_urdu_stem.yaml +4 -0
- lm-evaluation-harness/lm_eval/tasks/include/few_shot_en/Urdu/utils.py +30 -0
- lm-evaluation-harness/lm_eval/tasks/include/few_shot_en/Uzbek/_include_base_44_uzbek.yaml +11 -0
- lm-evaluation-harness/lm_eval/tasks/include/few_shot_en/Uzbek/_uzbek_few_shot_en_template_yaml +18 -0
- lm-evaluation-harness/lm_eval/tasks/include/few_shot_en/Uzbek/include_base_44_uzbek_arts_humanities.yaml +5 -0
- lm-evaluation-harness/lm_eval/tasks/include/few_shot_en/Uzbek/include_base_44_uzbek_medical_license.yaml +5 -0
- lm-evaluation-harness/lm_eval/tasks/include/few_shot_en/Uzbek/include_base_44_uzbek_social_science.yaml +5 -0
- lm-evaluation-harness/lm_eval/tasks/include/few_shot_en/Uzbek/include_base_44_uzbek_stem.yaml +4 -0
- lm-evaluation-harness/lm_eval/tasks/include/few_shot_en/Uzbek/utils.py +30 -0
- lm-evaluation-harness/lm_eval/tasks/include/few_shot_en/Vietnamese/_include_base_44_vietnamese.yaml +10 -0
- lm-evaluation-harness/lm_eval/tasks/include/few_shot_en/Vietnamese/_vietnamese_few_shot_en_template_yaml +18 -0
- lm-evaluation-harness/lm_eval/tasks/include/few_shot_en/Vietnamese/include_base_44_vietnamese_arts_humanities.yaml +5 -0
- lm-evaluation-harness/lm_eval/tasks/include/few_shot_en/Vietnamese/include_base_44_vietnamese_social_science.yaml +5 -0
- lm-evaluation-harness/lm_eval/tasks/include/few_shot_en/Vietnamese/include_base_44_vietnamese_stem.yaml +4 -0
- lm-evaluation-harness/lm_eval/tasks/include/few_shot_en/Vietnamese/utils.py +30 -0
- lm-evaluation-harness/lm_eval/tasks/include/few_shot_og/Albanian/_albanian_few_shot_og_template_yaml +18 -0
- lm-evaluation-harness/lm_eval/tasks/include/few_shot_og/Albanian/_include_base_44_albanian.yaml +12 -0
- lm-evaluation-harness/lm_eval/tasks/include/few_shot_og/Albanian/include_base_44_albanian_arts_humanities.yaml +5 -0
- lm-evaluation-harness/lm_eval/tasks/include/few_shot_og/Albanian/include_base_44_albanian_business_commerce.yaml +5 -0
- lm-evaluation-harness/lm_eval/tasks/include/few_shot_og/Albanian/include_base_44_albanian_health_oriented_education.yaml +5 -0
- lm-evaluation-harness/lm_eval/tasks/include/few_shot_og/Albanian/include_base_44_albanian_social_science.yaml +5 -0
- lm-evaluation-harness/lm_eval/tasks/include/few_shot_og/Albanian/include_base_44_albanian_stem.yaml +5 -0
lm-evaluation-harness/lm_eval/tasks/include/few_shot_en/Serbian/utils.py
ADDED
|
@@ -0,0 +1,30 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
from functools import partial
|
| 2 |
+
|
| 3 |
+
|
| 4 |
+
CATEGORIES = [
|
| 5 |
+
"Applied Science",
|
| 6 |
+
"Arts & Humanities",
|
| 7 |
+
"Business & Commerce",
|
| 8 |
+
"Driving License",
|
| 9 |
+
"General knowledge",
|
| 10 |
+
"Health oriented education",
|
| 11 |
+
"Marine License",
|
| 12 |
+
"Medical License",
|
| 13 |
+
"Professional certification",
|
| 14 |
+
"STEM",
|
| 15 |
+
"Social Science",
|
| 16 |
+
]
|
| 17 |
+
|
| 18 |
+
|
| 19 |
+
def process_docs(dataset, category):
|
| 20 |
+
return dataset.filter(lambda x: x["domain"] == category)
|
| 21 |
+
|
| 22 |
+
|
| 23 |
+
process_functions = {
|
| 24 |
+
f"process_{category.lower().replace(' & ', '_').replace(' ', '_')}": partial(
|
| 25 |
+
process_docs, category=category
|
| 26 |
+
)
|
| 27 |
+
for category in CATEGORIES
|
| 28 |
+
}
|
| 29 |
+
|
| 30 |
+
globals().update(process_functions)
|
lm-evaluation-harness/lm_eval/tasks/include/few_shot_en/Spanish/include_base_44_spanish_health_oriented_education.yaml
ADDED
|
@@ -0,0 +1,5 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
include: _spanish_few_shot_en_template_yaml
|
| 2 |
+
description: The following are multiple-choice questions (with answers) about Health
|
| 3 |
+
oriented education.
|
| 4 |
+
process_docs: !function 'utils.process_health_oriented_education'
|
| 5 |
+
task: include_base_44_spanish_few_shot_en_health_oriented_education
|
lm-evaluation-harness/lm_eval/tasks/include/few_shot_en/Spanish/include_base_44_spanish_stem.yaml
ADDED
|
@@ -0,0 +1,4 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
include: _spanish_few_shot_en_template_yaml
|
| 2 |
+
description: The following are multiple-choice questions (with answers) about STEM.
|
| 3 |
+
process_docs: !function 'utils.process_stem'
|
| 4 |
+
task: include_base_44_spanish_few_shot_en_stem
|
lm-evaluation-harness/lm_eval/tasks/include/few_shot_en/Spanish/utils.py
ADDED
|
@@ -0,0 +1,30 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
from functools import partial
|
| 2 |
+
|
| 3 |
+
|
| 4 |
+
CATEGORIES = [
|
| 5 |
+
"Applied Science",
|
| 6 |
+
"Arts & Humanities",
|
| 7 |
+
"Business & Commerce",
|
| 8 |
+
"Driving License",
|
| 9 |
+
"General knowledge",
|
| 10 |
+
"Health oriented education",
|
| 11 |
+
"Marine License",
|
| 12 |
+
"Medical License",
|
| 13 |
+
"Professional certification",
|
| 14 |
+
"STEM",
|
| 15 |
+
"Social Science",
|
| 16 |
+
]
|
| 17 |
+
|
| 18 |
+
|
| 19 |
+
def process_docs(dataset, category):
|
| 20 |
+
return dataset.filter(lambda x: x["domain"] == category)
|
| 21 |
+
|
| 22 |
+
|
| 23 |
+
process_functions = {
|
| 24 |
+
f"process_{category.lower().replace(' & ', '_').replace(' ', '_')}": partial(
|
| 25 |
+
process_docs, category=category
|
| 26 |
+
)
|
| 27 |
+
for category in CATEGORIES
|
| 28 |
+
}
|
| 29 |
+
|
| 30 |
+
globals().update(process_functions)
|
lm-evaluation-harness/lm_eval/tasks/include/few_shot_en/Tagalog/_tagalog_few_shot_en_template_yaml
ADDED
|
@@ -0,0 +1,18 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
dataset_path: CohereForAI/include-base-44
|
| 2 |
+
dataset_name: Tagalog
|
| 3 |
+
test_split: test
|
| 4 |
+
output_type: multiple_choice
|
| 5 |
+
doc_to_text: "{{question.strip()}}\nA. {{option_a}}\nB. {{option_b}}\nC. {{option_c}}\n
|
| 6 |
+
D. {{option_d}}\nAnswer:"
|
| 7 |
+
doc_to_choice:
|
| 8 |
+
- A
|
| 9 |
+
- B
|
| 10 |
+
- C
|
| 11 |
+
- D
|
| 12 |
+
doc_to_target: answer
|
| 13 |
+
metric_list:
|
| 14 |
+
- metric: acc
|
| 15 |
+
aggregation: mean
|
| 16 |
+
higher_is_better: true
|
| 17 |
+
metadata:
|
| 18 |
+
version: 0.0
|
lm-evaluation-harness/lm_eval/tasks/include/few_shot_en/Tagalog/include_base_44_tagalog_arts_humanities.yaml
ADDED
|
@@ -0,0 +1,5 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
include: _tagalog_few_shot_en_template_yaml
|
| 2 |
+
description: The following are multiple-choice questions (with answers) about Arts
|
| 3 |
+
& Humanities.
|
| 4 |
+
process_docs: !function 'utils.process_arts_humanities'
|
| 5 |
+
task: include_base_44_tagalog_few_shot_en_arts_humanities
|
lm-evaluation-harness/lm_eval/tasks/include/few_shot_en/Tamil/_include_base_44_tamil.yaml
ADDED
|
@@ -0,0 +1,9 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
group: include_base_44_tamil
|
| 2 |
+
task:
|
| 3 |
+
- include_base_44_tamil_few_shot_en_stem
|
| 4 |
+
- include_base_44_tamil_few_shot_en_general_knowledge
|
| 5 |
+
aggregate_metric_list:
|
| 6 |
+
- metric: acc
|
| 7 |
+
weight_by_size: true
|
| 8 |
+
metadata:
|
| 9 |
+
version: 0.0
|
lm-evaluation-harness/lm_eval/tasks/include/few_shot_en/Tamil/include_base_44_tamil_general_knowledge.yaml
ADDED
|
@@ -0,0 +1,5 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
include: _tamil_few_shot_en_template_yaml
|
| 2 |
+
description: The following are multiple-choice questions (with answers) about General
|
| 3 |
+
knowledge.
|
| 4 |
+
process_docs: !function 'utils.process_general_knowledge'
|
| 5 |
+
task: include_base_44_tamil_few_shot_en_general_knowledge
|
lm-evaluation-harness/lm_eval/tasks/include/few_shot_en/Tamil/include_base_44_tamil_stem.yaml
ADDED
|
@@ -0,0 +1,4 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
include: _tamil_few_shot_en_template_yaml
|
| 2 |
+
description: The following are multiple-choice questions (with answers) about STEM.
|
| 3 |
+
process_docs: !function 'utils.process_stem'
|
| 4 |
+
task: include_base_44_tamil_few_shot_en_stem
|
lm-evaluation-harness/lm_eval/tasks/include/few_shot_en/Telugu/_include_base_44_telugu.yaml
ADDED
|
@@ -0,0 +1,11 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
group: include_base_44_telugu
|
| 2 |
+
task:
|
| 3 |
+
- include_base_44_telugu_few_shot_en_arts_humanities
|
| 4 |
+
- include_base_44_telugu_few_shot_en_applied_science
|
| 5 |
+
- include_base_44_telugu_few_shot_en_stem
|
| 6 |
+
- include_base_44_telugu_few_shot_en_social_science
|
| 7 |
+
aggregate_metric_list:
|
| 8 |
+
- metric: acc
|
| 9 |
+
weight_by_size: true
|
| 10 |
+
metadata:
|
| 11 |
+
version: 0.0
|
lm-evaluation-harness/lm_eval/tasks/include/few_shot_en/Telugu/include_base_44_telugu_arts_humanities.yaml
ADDED
|
@@ -0,0 +1,5 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
include: _telugu_few_shot_en_template_yaml
|
| 2 |
+
description: The following are multiple-choice questions (with answers) about Arts
|
| 3 |
+
& Humanities.
|
| 4 |
+
process_docs: !function 'utils.process_arts_humanities'
|
| 5 |
+
task: include_base_44_telugu_few_shot_en_arts_humanities
|
lm-evaluation-harness/lm_eval/tasks/include/few_shot_en/Telugu/include_base_44_telugu_social_science.yaml
ADDED
|
@@ -0,0 +1,5 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
include: _telugu_few_shot_en_template_yaml
|
| 2 |
+
description: The following are multiple-choice questions (with answers) about Social
|
| 3 |
+
Science.
|
| 4 |
+
process_docs: !function 'utils.process_social_science'
|
| 5 |
+
task: include_base_44_telugu_few_shot_en_social_science
|
lm-evaluation-harness/lm_eval/tasks/include/few_shot_en/Telugu/utils.py
ADDED
|
@@ -0,0 +1,30 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
from functools import partial
|
| 2 |
+
|
| 3 |
+
|
| 4 |
+
CATEGORIES = [
|
| 5 |
+
"Applied Science",
|
| 6 |
+
"Arts & Humanities",
|
| 7 |
+
"Business & Commerce",
|
| 8 |
+
"Driving License",
|
| 9 |
+
"General knowledge",
|
| 10 |
+
"Health oriented education",
|
| 11 |
+
"Marine License",
|
| 12 |
+
"Medical License",
|
| 13 |
+
"Professional certification",
|
| 14 |
+
"STEM",
|
| 15 |
+
"Social Science",
|
| 16 |
+
]
|
| 17 |
+
|
| 18 |
+
|
| 19 |
+
def process_docs(dataset, category):
|
| 20 |
+
return dataset.filter(lambda x: x["domain"] == category)
|
| 21 |
+
|
| 22 |
+
|
| 23 |
+
process_functions = {
|
| 24 |
+
f"process_{category.lower().replace(' & ', '_').replace(' ', '_')}": partial(
|
| 25 |
+
process_docs, category=category
|
| 26 |
+
)
|
| 27 |
+
for category in CATEGORIES
|
| 28 |
+
}
|
| 29 |
+
|
| 30 |
+
globals().update(process_functions)
|
lm-evaluation-harness/lm_eval/tasks/include/few_shot_en/Turkish/_include_base_44_turkish.yaml
ADDED
|
@@ -0,0 +1,11 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
group: include_base_44_turkish
|
| 2 |
+
task:
|
| 3 |
+
- include_base_44_turkish_few_shot_en_business_commerce
|
| 4 |
+
- include_base_44_turkish_few_shot_en_stem
|
| 5 |
+
- include_base_44_turkish_few_shot_en_arts_humanities
|
| 6 |
+
- include_base_44_turkish_few_shot_en_social_science
|
| 7 |
+
aggregate_metric_list:
|
| 8 |
+
- metric: acc
|
| 9 |
+
weight_by_size: true
|
| 10 |
+
metadata:
|
| 11 |
+
version: 0.0
|
lm-evaluation-harness/lm_eval/tasks/include/few_shot_en/Turkish/include_base_44_turkish_business_commerce.yaml
ADDED
|
@@ -0,0 +1,5 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
include: _turkish_few_shot_en_template_yaml
|
| 2 |
+
description: The following are multiple-choice questions (with answers) about Business
|
| 3 |
+
& Commerce.
|
| 4 |
+
process_docs: !function 'utils.process_business_commerce'
|
| 5 |
+
task: include_base_44_turkish_few_shot_en_business_commerce
|
lm-evaluation-harness/lm_eval/tasks/include/few_shot_en/Turkish/include_base_44_turkish_social_science.yaml
ADDED
|
@@ -0,0 +1,5 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
include: _turkish_few_shot_en_template_yaml
|
| 2 |
+
description: The following are multiple-choice questions (with answers) about Social
|
| 3 |
+
Science.
|
| 4 |
+
process_docs: !function 'utils.process_social_science'
|
| 5 |
+
task: include_base_44_turkish_few_shot_en_social_science
|
lm-evaluation-harness/lm_eval/tasks/include/few_shot_en/Turkish/include_base_44_turkish_stem.yaml
ADDED
|
@@ -0,0 +1,4 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
include: _turkish_few_shot_en_template_yaml
|
| 2 |
+
description: The following are multiple-choice questions (with answers) about STEM.
|
| 3 |
+
process_docs: !function 'utils.process_stem'
|
| 4 |
+
task: include_base_44_turkish_few_shot_en_stem
|
lm-evaluation-harness/lm_eval/tasks/include/few_shot_en/Turkish/utils.py
ADDED
|
@@ -0,0 +1,30 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
from functools import partial
|
| 2 |
+
|
| 3 |
+
|
| 4 |
+
CATEGORIES = [
|
| 5 |
+
"Applied Science",
|
| 6 |
+
"Arts & Humanities",
|
| 7 |
+
"Business & Commerce",
|
| 8 |
+
"Driving License",
|
| 9 |
+
"General knowledge",
|
| 10 |
+
"Health oriented education",
|
| 11 |
+
"Marine License",
|
| 12 |
+
"Medical License",
|
| 13 |
+
"Professional certification",
|
| 14 |
+
"STEM",
|
| 15 |
+
"Social Science",
|
| 16 |
+
]
|
| 17 |
+
|
| 18 |
+
|
| 19 |
+
def process_docs(dataset, category):
|
| 20 |
+
return dataset.filter(lambda x: x["domain"] == category)
|
| 21 |
+
|
| 22 |
+
|
| 23 |
+
process_functions = {
|
| 24 |
+
f"process_{category.lower().replace(' & ', '_').replace(' ', '_')}": partial(
|
| 25 |
+
process_docs, category=category
|
| 26 |
+
)
|
| 27 |
+
for category in CATEGORIES
|
| 28 |
+
}
|
| 29 |
+
|
| 30 |
+
globals().update(process_functions)
|
lm-evaluation-harness/lm_eval/tasks/include/few_shot_en/Ukrainian/_include_base_44_ukrainian.yaml
ADDED
|
@@ -0,0 +1,10 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
group: include_base_44_ukrainian
|
| 2 |
+
task:
|
| 3 |
+
- include_base_44_ukrainian_few_shot_en_arts_humanities
|
| 4 |
+
- include_base_44_ukrainian_few_shot_en_social_science
|
| 5 |
+
- include_base_44_ukrainian_few_shot_en_stem
|
| 6 |
+
aggregate_metric_list:
|
| 7 |
+
- metric: acc
|
| 8 |
+
weight_by_size: true
|
| 9 |
+
metadata:
|
| 10 |
+
version: 0.0
|
lm-evaluation-harness/lm_eval/tasks/include/few_shot_en/Ukrainian/_ukrainian_few_shot_en_template_yaml
ADDED
|
@@ -0,0 +1,18 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
dataset_path: CohereForAI/include-base-44
|
| 2 |
+
dataset_name: Ukrainian
|
| 3 |
+
test_split: test
|
| 4 |
+
output_type: multiple_choice
|
| 5 |
+
doc_to_text: "{{question.strip()}}\nA. {{option_a}}\nB. {{option_b}}\nC. {{option_c}}\n
|
| 6 |
+
D. {{option_d}}\nAnswer:"
|
| 7 |
+
doc_to_choice:
|
| 8 |
+
- A
|
| 9 |
+
- B
|
| 10 |
+
- C
|
| 11 |
+
- D
|
| 12 |
+
doc_to_target: answer
|
| 13 |
+
metric_list:
|
| 14 |
+
- metric: acc
|
| 15 |
+
aggregation: mean
|
| 16 |
+
higher_is_better: true
|
| 17 |
+
metadata:
|
| 18 |
+
version: 0.0
|
lm-evaluation-harness/lm_eval/tasks/include/few_shot_en/Ukrainian/include_base_44_ukrainian_arts_humanities.yaml
ADDED
|
@@ -0,0 +1,5 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
include: _ukrainian_few_shot_en_template_yaml
|
| 2 |
+
description: The following are multiple-choice questions (with answers) about Arts
|
| 3 |
+
& Humanities.
|
| 4 |
+
process_docs: !function 'utils.process_arts_humanities'
|
| 5 |
+
task: include_base_44_ukrainian_few_shot_en_arts_humanities
|
lm-evaluation-harness/lm_eval/tasks/include/few_shot_en/Ukrainian/include_base_44_ukrainian_social_science.yaml
ADDED
|
@@ -0,0 +1,5 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
include: _ukrainian_few_shot_en_template_yaml
|
| 2 |
+
description: The following are multiple-choice questions (with answers) about Social
|
| 3 |
+
Science.
|
| 4 |
+
process_docs: !function 'utils.process_social_science'
|
| 5 |
+
task: include_base_44_ukrainian_few_shot_en_social_science
|
lm-evaluation-harness/lm_eval/tasks/include/few_shot_en/Ukrainian/include_base_44_ukrainian_stem.yaml
ADDED
|
@@ -0,0 +1,4 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
include: _ukrainian_few_shot_en_template_yaml
|
| 2 |
+
description: The following are multiple-choice questions (with answers) about STEM.
|
| 3 |
+
process_docs: !function 'utils.process_stem'
|
| 4 |
+
task: include_base_44_ukrainian_few_shot_en_stem
|
lm-evaluation-harness/lm_eval/tasks/include/few_shot_en/Ukrainian/utils.py
ADDED
|
@@ -0,0 +1,30 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
from functools import partial
|
| 2 |
+
|
| 3 |
+
|
| 4 |
+
CATEGORIES = [
|
| 5 |
+
"Applied Science",
|
| 6 |
+
"Arts & Humanities",
|
| 7 |
+
"Business & Commerce",
|
| 8 |
+
"Driving License",
|
| 9 |
+
"General knowledge",
|
| 10 |
+
"Health oriented education",
|
| 11 |
+
"Marine License",
|
| 12 |
+
"Medical License",
|
| 13 |
+
"Professional certification",
|
| 14 |
+
"STEM",
|
| 15 |
+
"Social Science",
|
| 16 |
+
]
|
| 17 |
+
|
| 18 |
+
|
| 19 |
+
def process_docs(dataset, category):
|
| 20 |
+
return dataset.filter(lambda x: x["domain"] == category)
|
| 21 |
+
|
| 22 |
+
|
| 23 |
+
process_functions = {
|
| 24 |
+
f"process_{category.lower().replace(' & ', '_').replace(' ', '_')}": partial(
|
| 25 |
+
process_docs, category=category
|
| 26 |
+
)
|
| 27 |
+
for category in CATEGORIES
|
| 28 |
+
}
|
| 29 |
+
|
| 30 |
+
globals().update(process_functions)
|
lm-evaluation-harness/lm_eval/tasks/include/few_shot_en/Urdu/_include_base_44_urdu.yaml
ADDED
|
@@ -0,0 +1,10 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
group: include_base_44_urdu
|
| 2 |
+
task:
|
| 3 |
+
- include_base_44_urdu_few_shot_en_stem
|
| 4 |
+
- include_base_44_urdu_few_shot_en_health_oriented_education
|
| 5 |
+
- include_base_44_urdu_few_shot_en_arts_humanities
|
| 6 |
+
aggregate_metric_list:
|
| 7 |
+
- metric: acc
|
| 8 |
+
weight_by_size: true
|
| 9 |
+
metadata:
|
| 10 |
+
version: 0.0
|
lm-evaluation-harness/lm_eval/tasks/include/few_shot_en/Urdu/_urdu_few_shot_en_template_yaml
ADDED
|
@@ -0,0 +1,18 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
dataset_path: CohereForAI/include-base-44
|
| 2 |
+
dataset_name: Urdu
|
| 3 |
+
test_split: test
|
| 4 |
+
output_type: multiple_choice
|
| 5 |
+
doc_to_text: "{{question.strip()}}\nA. {{option_a}}\nB. {{option_b}}\nC. {{option_c}}\n
|
| 6 |
+
D. {{option_d}}\nAnswer:"
|
| 7 |
+
doc_to_choice:
|
| 8 |
+
- A
|
| 9 |
+
- B
|
| 10 |
+
- C
|
| 11 |
+
- D
|
| 12 |
+
doc_to_target: answer
|
| 13 |
+
metric_list:
|
| 14 |
+
- metric: acc
|
| 15 |
+
aggregation: mean
|
| 16 |
+
higher_is_better: true
|
| 17 |
+
metadata:
|
| 18 |
+
version: 0.0
|
lm-evaluation-harness/lm_eval/tasks/include/few_shot_en/Urdu/include_base_44_urdu_arts_humanities.yaml
ADDED
|
@@ -0,0 +1,5 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
include: _urdu_few_shot_en_template_yaml
|
| 2 |
+
description: The following are multiple-choice questions (with answers) about Arts
|
| 3 |
+
& Humanities.
|
| 4 |
+
process_docs: !function 'utils.process_arts_humanities'
|
| 5 |
+
task: include_base_44_urdu_few_shot_en_arts_humanities
|
lm-evaluation-harness/lm_eval/tasks/include/few_shot_en/Urdu/include_base_44_urdu_health_oriented_education.yaml
ADDED
|
@@ -0,0 +1,5 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
include: _urdu_few_shot_en_template_yaml
|
| 2 |
+
description: The following are multiple-choice questions (with answers) about Health
|
| 3 |
+
oriented education.
|
| 4 |
+
process_docs: !function 'utils.process_health_oriented_education'
|
| 5 |
+
task: include_base_44_urdu_few_shot_en_health_oriented_education
|
lm-evaluation-harness/lm_eval/tasks/include/few_shot_en/Urdu/include_base_44_urdu_stem.yaml
ADDED
|
@@ -0,0 +1,4 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
include: _urdu_few_shot_en_template_yaml
|
| 2 |
+
description: The following are multiple-choice questions (with answers) about STEM.
|
| 3 |
+
process_docs: !function 'utils.process_stem'
|
| 4 |
+
task: include_base_44_urdu_few_shot_en_stem
|
lm-evaluation-harness/lm_eval/tasks/include/few_shot_en/Urdu/utils.py
ADDED
|
@@ -0,0 +1,30 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
from functools import partial
|
| 2 |
+
|
| 3 |
+
|
| 4 |
+
CATEGORIES = [
|
| 5 |
+
"Applied Science",
|
| 6 |
+
"Arts & Humanities",
|
| 7 |
+
"Business & Commerce",
|
| 8 |
+
"Driving License",
|
| 9 |
+
"General knowledge",
|
| 10 |
+
"Health oriented education",
|
| 11 |
+
"Marine License",
|
| 12 |
+
"Medical License",
|
| 13 |
+
"Professional certification",
|
| 14 |
+
"STEM",
|
| 15 |
+
"Social Science",
|
| 16 |
+
]
|
| 17 |
+
|
| 18 |
+
|
| 19 |
+
def process_docs(dataset, category):
|
| 20 |
+
return dataset.filter(lambda x: x["domain"] == category)
|
| 21 |
+
|
| 22 |
+
|
| 23 |
+
process_functions = {
|
| 24 |
+
f"process_{category.lower().replace(' & ', '_').replace(' ', '_')}": partial(
|
| 25 |
+
process_docs, category=category
|
| 26 |
+
)
|
| 27 |
+
for category in CATEGORIES
|
| 28 |
+
}
|
| 29 |
+
|
| 30 |
+
globals().update(process_functions)
|
lm-evaluation-harness/lm_eval/tasks/include/few_shot_en/Uzbek/_include_base_44_uzbek.yaml
ADDED
|
@@ -0,0 +1,11 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
group: include_base_44_uzbek
|
| 2 |
+
task:
|
| 3 |
+
- include_base_44_uzbek_few_shot_en_arts_humanities
|
| 4 |
+
- include_base_44_uzbek_few_shot_en_medical_license
|
| 5 |
+
- include_base_44_uzbek_few_shot_en_social_science
|
| 6 |
+
- include_base_44_uzbek_few_shot_en_stem
|
| 7 |
+
aggregate_metric_list:
|
| 8 |
+
- metric: acc
|
| 9 |
+
weight_by_size: true
|
| 10 |
+
metadata:
|
| 11 |
+
version: 0.0
|
lm-evaluation-harness/lm_eval/tasks/include/few_shot_en/Uzbek/_uzbek_few_shot_en_template_yaml
ADDED
|
@@ -0,0 +1,18 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
dataset_path: CohereForAI/include-base-44
|
| 2 |
+
dataset_name: Uzbek
|
| 3 |
+
test_split: test
|
| 4 |
+
output_type: multiple_choice
|
| 5 |
+
doc_to_text: "{{question.strip()}}\nA. {{option_a}}\nB. {{option_b}}\nC. {{option_c}}\n
|
| 6 |
+
D. {{option_d}}\nAnswer:"
|
| 7 |
+
doc_to_choice:
|
| 8 |
+
- A
|
| 9 |
+
- B
|
| 10 |
+
- C
|
| 11 |
+
- D
|
| 12 |
+
doc_to_target: answer
|
| 13 |
+
metric_list:
|
| 14 |
+
- metric: acc
|
| 15 |
+
aggregation: mean
|
| 16 |
+
higher_is_better: true
|
| 17 |
+
metadata:
|
| 18 |
+
version: 0.0
|
lm-evaluation-harness/lm_eval/tasks/include/few_shot_en/Uzbek/include_base_44_uzbek_arts_humanities.yaml
ADDED
|
@@ -0,0 +1,5 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
include: _uzbek_few_shot_en_template_yaml
|
| 2 |
+
description: The following are multiple-choice questions (with answers) about Arts
|
| 3 |
+
& Humanities.
|
| 4 |
+
process_docs: !function 'utils.process_arts_humanities'
|
| 5 |
+
task: include_base_44_uzbek_few_shot_en_arts_humanities
|
lm-evaluation-harness/lm_eval/tasks/include/few_shot_en/Uzbek/include_base_44_uzbek_medical_license.yaml
ADDED
|
@@ -0,0 +1,5 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
include: _uzbek_few_shot_en_template_yaml
|
| 2 |
+
description: The following are multiple-choice questions (with answers) about Medical
|
| 3 |
+
License.
|
| 4 |
+
process_docs: !function 'utils.process_medical_license'
|
| 5 |
+
task: include_base_44_uzbek_few_shot_en_medical_license
|
lm-evaluation-harness/lm_eval/tasks/include/few_shot_en/Uzbek/include_base_44_uzbek_social_science.yaml
ADDED
|
@@ -0,0 +1,5 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
include: _uzbek_few_shot_en_template_yaml
|
| 2 |
+
description: The following are multiple-choice questions (with answers) about Social
|
| 3 |
+
Science.
|
| 4 |
+
process_docs: !function 'utils.process_social_science'
|
| 5 |
+
task: include_base_44_uzbek_few_shot_en_social_science
|
lm-evaluation-harness/lm_eval/tasks/include/few_shot_en/Uzbek/include_base_44_uzbek_stem.yaml
ADDED
|
@@ -0,0 +1,4 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
include: _uzbek_few_shot_en_template_yaml
|
| 2 |
+
description: The following are multiple-choice questions (with answers) about STEM.
|
| 3 |
+
process_docs: !function 'utils.process_stem'
|
| 4 |
+
task: include_base_44_uzbek_few_shot_en_stem
|
lm-evaluation-harness/lm_eval/tasks/include/few_shot_en/Uzbek/utils.py
ADDED
|
@@ -0,0 +1,30 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
from functools import partial
|
| 2 |
+
|
| 3 |
+
|
| 4 |
+
CATEGORIES = [
|
| 5 |
+
"Applied Science",
|
| 6 |
+
"Arts & Humanities",
|
| 7 |
+
"Business & Commerce",
|
| 8 |
+
"Driving License",
|
| 9 |
+
"General knowledge",
|
| 10 |
+
"Health oriented education",
|
| 11 |
+
"Marine License",
|
| 12 |
+
"Medical License",
|
| 13 |
+
"Professional certification",
|
| 14 |
+
"STEM",
|
| 15 |
+
"Social Science",
|
| 16 |
+
]
|
| 17 |
+
|
| 18 |
+
|
| 19 |
+
def process_docs(dataset, category):
|
| 20 |
+
return dataset.filter(lambda x: x["domain"] == category)
|
| 21 |
+
|
| 22 |
+
|
| 23 |
+
process_functions = {
|
| 24 |
+
f"process_{category.lower().replace(' & ', '_').replace(' ', '_')}": partial(
|
| 25 |
+
process_docs, category=category
|
| 26 |
+
)
|
| 27 |
+
for category in CATEGORIES
|
| 28 |
+
}
|
| 29 |
+
|
| 30 |
+
globals().update(process_functions)
|
lm-evaluation-harness/lm_eval/tasks/include/few_shot_en/Vietnamese/_include_base_44_vietnamese.yaml
ADDED
|
@@ -0,0 +1,10 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
group: include_base_44_vietnamese
|
| 2 |
+
task:
|
| 3 |
+
- include_base_44_vietnamese_few_shot_en_stem
|
| 4 |
+
- include_base_44_vietnamese_few_shot_en_arts_humanities
|
| 5 |
+
- include_base_44_vietnamese_few_shot_en_social_science
|
| 6 |
+
aggregate_metric_list:
|
| 7 |
+
- metric: acc
|
| 8 |
+
weight_by_size: true
|
| 9 |
+
metadata:
|
| 10 |
+
version: 0.0
|
lm-evaluation-harness/lm_eval/tasks/include/few_shot_en/Vietnamese/_vietnamese_few_shot_en_template_yaml
ADDED
|
@@ -0,0 +1,18 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
dataset_path: CohereForAI/include-base-44
|
| 2 |
+
dataset_name: Vietnamese
|
| 3 |
+
test_split: test
|
| 4 |
+
output_type: multiple_choice
|
| 5 |
+
doc_to_text: "{{question.strip()}}\nA. {{option_a}}\nB. {{option_b}}\nC. {{option_c}}\n
|
| 6 |
+
D. {{option_d}}\nAnswer:"
|
| 7 |
+
doc_to_choice:
|
| 8 |
+
- A
|
| 9 |
+
- B
|
| 10 |
+
- C
|
| 11 |
+
- D
|
| 12 |
+
doc_to_target: answer
|
| 13 |
+
metric_list:
|
| 14 |
+
- metric: acc
|
| 15 |
+
aggregation: mean
|
| 16 |
+
higher_is_better: true
|
| 17 |
+
metadata:
|
| 18 |
+
version: 0.0
|
lm-evaluation-harness/lm_eval/tasks/include/few_shot_en/Vietnamese/include_base_44_vietnamese_arts_humanities.yaml
ADDED
|
@@ -0,0 +1,5 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
include: _vietnamese_few_shot_en_template_yaml
|
| 2 |
+
description: The following are multiple-choice questions (with answers) about Arts
|
| 3 |
+
& Humanities.
|
| 4 |
+
process_docs: !function 'utils.process_arts_humanities'
|
| 5 |
+
task: include_base_44_vietnamese_few_shot_en_arts_humanities
|
lm-evaluation-harness/lm_eval/tasks/include/few_shot_en/Vietnamese/include_base_44_vietnamese_social_science.yaml
ADDED
|
@@ -0,0 +1,5 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
include: _vietnamese_few_shot_en_template_yaml
|
| 2 |
+
description: The following are multiple-choice questions (with answers) about Social
|
| 3 |
+
Science.
|
| 4 |
+
process_docs: !function 'utils.process_social_science'
|
| 5 |
+
task: include_base_44_vietnamese_few_shot_en_social_science
|
lm-evaluation-harness/lm_eval/tasks/include/few_shot_en/Vietnamese/include_base_44_vietnamese_stem.yaml
ADDED
|
@@ -0,0 +1,4 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
include: _vietnamese_few_shot_en_template_yaml
|
| 2 |
+
description: The following are multiple-choice questions (with answers) about STEM.
|
| 3 |
+
process_docs: !function 'utils.process_stem'
|
| 4 |
+
task: include_base_44_vietnamese_few_shot_en_stem
|
lm-evaluation-harness/lm_eval/tasks/include/few_shot_en/Vietnamese/utils.py
ADDED
|
@@ -0,0 +1,30 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
from functools import partial
|
| 2 |
+
|
| 3 |
+
|
| 4 |
+
CATEGORIES = [
|
| 5 |
+
"Applied Science",
|
| 6 |
+
"Arts & Humanities",
|
| 7 |
+
"Business & Commerce",
|
| 8 |
+
"Driving License",
|
| 9 |
+
"General knowledge",
|
| 10 |
+
"Health oriented education",
|
| 11 |
+
"Marine License",
|
| 12 |
+
"Medical License",
|
| 13 |
+
"Professional certification",
|
| 14 |
+
"STEM",
|
| 15 |
+
"Social Science",
|
| 16 |
+
]
|
| 17 |
+
|
| 18 |
+
|
| 19 |
+
def process_docs(dataset, category):
|
| 20 |
+
return dataset.filter(lambda x: x["domain"] == category)
|
| 21 |
+
|
| 22 |
+
|
| 23 |
+
process_functions = {
|
| 24 |
+
f"process_{category.lower().replace(' & ', '_').replace(' ', '_')}": partial(
|
| 25 |
+
process_docs, category=category
|
| 26 |
+
)
|
| 27 |
+
for category in CATEGORIES
|
| 28 |
+
}
|
| 29 |
+
|
| 30 |
+
globals().update(process_functions)
|
lm-evaluation-harness/lm_eval/tasks/include/few_shot_og/Albanian/_albanian_few_shot_og_template_yaml
ADDED
|
@@ -0,0 +1,18 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
dataset_path: CohereForAI/include-base-44
|
| 2 |
+
dataset_name: Albanian
|
| 3 |
+
test_split: test
|
| 4 |
+
output_type: multiple_choice
|
| 5 |
+
doc_to_text: "{{question.strip()}}\nA. {{option_a}}\nB. {{option_b}}\nC. {{option_c}}\n
|
| 6 |
+
D. {{option_d}}\nAnswer:"
|
| 7 |
+
doc_to_choice:
|
| 8 |
+
- A
|
| 9 |
+
- B
|
| 10 |
+
- C
|
| 11 |
+
- D
|
| 12 |
+
doc_to_target: answer
|
| 13 |
+
metric_list:
|
| 14 |
+
- metric: acc
|
| 15 |
+
aggregation: mean
|
| 16 |
+
higher_is_better: true
|
| 17 |
+
metadata:
|
| 18 |
+
version: 0.0
|
lm-evaluation-harness/lm_eval/tasks/include/few_shot_og/Albanian/_include_base_44_albanian.yaml
ADDED
|
@@ -0,0 +1,12 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
group: include_base_44_albanian
|
| 2 |
+
task:
|
| 3 |
+
- include_base_44_albanian_few_shot_og_arts_humanities
|
| 4 |
+
- include_base_44_albanian_few_shot_og_stem
|
| 5 |
+
- include_base_44_albanian_few_shot_og_business_commerce
|
| 6 |
+
- include_base_44_albanian_few_shot_og_social_science
|
| 7 |
+
- include_base_44_albanian_few_shot_og_health_oriented_education
|
| 8 |
+
aggregate_metric_list:
|
| 9 |
+
- metric: acc
|
| 10 |
+
weight_by_size: true
|
| 11 |
+
metadata:
|
| 12 |
+
version: 0.0
|
lm-evaluation-harness/lm_eval/tasks/include/few_shot_og/Albanian/include_base_44_albanian_arts_humanities.yaml
ADDED
|
@@ -0,0 +1,5 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
include: _albanian_few_shot_og_template_yaml
|
| 2 |
+
description: Më poshtë janë pyetjet me zgjedhje të shumëfishta (me përgjigje) rreth
|
| 3 |
+
Arts & Humanities.
|
| 4 |
+
process_docs: !function 'utils.process_arts_humanities'
|
| 5 |
+
task: include_base_44_albanian_few_shot_og_arts_humanities
|
lm-evaluation-harness/lm_eval/tasks/include/few_shot_og/Albanian/include_base_44_albanian_business_commerce.yaml
ADDED
|
@@ -0,0 +1,5 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
include: _albanian_few_shot_og_template_yaml
|
| 2 |
+
description: Më poshtë janë pyetjet me zgjedhje të shumëfishta (me përgjigje) rreth
|
| 3 |
+
Business & Commerce.
|
| 4 |
+
process_docs: !function 'utils.process_business_commerce'
|
| 5 |
+
task: include_base_44_albanian_few_shot_og_business_commerce
|
lm-evaluation-harness/lm_eval/tasks/include/few_shot_og/Albanian/include_base_44_albanian_health_oriented_education.yaml
ADDED
|
@@ -0,0 +1,5 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
include: _albanian_few_shot_og_template_yaml
|
| 2 |
+
description: Më poshtë janë pyetjet me zgjedhje të shumëfishta (me përgjigje) rreth
|
| 3 |
+
Health oriented education.
|
| 4 |
+
process_docs: !function 'utils.process_health_oriented_education'
|
| 5 |
+
task: include_base_44_albanian_few_shot_og_health_oriented_education
|
lm-evaluation-harness/lm_eval/tasks/include/few_shot_og/Albanian/include_base_44_albanian_social_science.yaml
ADDED
|
@@ -0,0 +1,5 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
include: _albanian_few_shot_og_template_yaml
|
| 2 |
+
description: Më poshtë janë pyetjet me zgjedhje të shumëfishta (me përgjigje) rreth
|
| 3 |
+
Social Science.
|
| 4 |
+
process_docs: !function 'utils.process_social_science'
|
| 5 |
+
task: include_base_44_albanian_few_shot_og_social_science
|
lm-evaluation-harness/lm_eval/tasks/include/few_shot_og/Albanian/include_base_44_albanian_stem.yaml
ADDED
|
@@ -0,0 +1,5 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
include: _albanian_few_shot_og_template_yaml
|
| 2 |
+
description: Më poshtë janë pyetjet me zgjedhje të shumëfishta (me përgjigje) rreth
|
| 3 |
+
STEM.
|
| 4 |
+
process_docs: !function 'utils.process_stem'
|
| 5 |
+
task: include_base_44_albanian_few_shot_og_stem
|