Add Open Arabic LLM Leaderboard Benchmarks (Full and Light Version) (#2232)

* arabic leaferboard yaml file is added

* arabic toxigen is implemented

* Dataset library is imported

* arabic sciq is added

* util file of arabic toxigen is updated

* arabic race is added

* arabic piqa is implemented

* arabic open qa is added

* arabic copa is implemented

* arabic boolq ia added

* arabic arc easy is added

* arabic arc challenge is added

* arabic exams benchmark is implemented

* arabic hellaswag is added

* arabic leaderboard yaml file metrics are updated

* arabic mmlu benchmarks are added

* arabic mmlu group yaml file is updated

* alghafa benchmarks are added

* acva benchmarks are added

* acva utils.py is updated

* light version of arabic leaderboard benchmarks are added

* bugs fixed

* bug fixed

* bug fixed

* bug fixed

* bug fixed

* bug fixed

* library import bug is fixed

* doc to target updated

* bash file is deleted

* results folder is deleted

* leaderboard groups are added

* full arabic leaderboard groups are added, plus some bug fixes to the light version

* Create README.md

README.md for arabic_leaderboard_complete

* Create README.md

README.md for arabic_leaderboard_light

* Delete lm_eval/tasks/arabic_leaderboard directory

* Update README.md

* Update README.md

adding the Arabic leaderboards to the library

* Update README.md

10% of the training set

* Update README.md

10% of the training set

* revert .gitignore to prev version

* Update lm_eval/tasks/README.md
Co-authored-by: Hailey Schoelkopf <65563625+haileyschoelkopf@users.noreply.github.com>

* updated main README.md

* Update lm_eval/tasks/README.md

* specify machine translated benchmarks (complete)

* specify machine translated benchmarks (light version)

* add alghafa to the related task names (complete and light)

* add 'acva' to the related task names (complete and light)

* add 'arabic_leaderboard' to all the groups (complete and light)

* all dataset - not a random sample

* added more accurate details to the readme file

* added mt_mmlu from okapi

* Update lm_eval/tasks/README.md
Co-authored-by: Hailey Schoelkopf <65563625+haileyschoelkopf@users.noreply.github.com>

* Update lm_eval/tasks/README.md
Co-authored-by: Hailey Schoelkopf <65563625+haileyschoelkopf@users.noreply.github.com>

* updated mt_mmlu readme

* renaming 'alghafa' full and light

* renaming 'arabic_mmlu' light and full

* renaming 'acva' full and light

* update readme and standardize dir/file names

* running pre-commit

---------
Co-authored-by: shahrzads <sayehban@ualberta.ca>
Co-authored-by: shahrzads <56282669+shahrzads@users.noreply.github.com>
Co-authored-by: Hailey Schoelkopf <65563625+haileyschoelkopf@users.noreply.github.com>

lm_eval/tasks/arabic_leaderboard_light/arabic_leaderboard_arabic_mt_mmlu_light/utils.py

+import datasets
+import numpy as np
+
+
+def process_docs(dataset: datasets.Dataset):
+    def _process_doc(doc):
+        question = doc["query"]
+        answer_index = int(doc["label"])
+        # Dynamically determining the choices by excluding '__few_shots', 'query' and 'label'
+        choices_keys = [
+            key for key in doc.keys() if key not in ["query", "label", "__few_shots"]
+        ]
+        choices = [doc[key] for key in choices_keys]
+
+        instruction = "الأسئلة التالية هي أسئلة متعددة الإختيارات مع الجواب الصحيح\n\n"
+        query = f"{instruction}السؤال: {question}\n"
+        for index, choice in enumerate(choices):
+            query += f"{index}) {choice}\n"
+        query += "الإجابة:"
+
+        return {"query": query, "choices": choices, "gold": answer_index}
+
+    return dataset.map(_process_doc)

lm_eval/tasks/arabic_leaderboard_light/arabic_leaderboard_arabic_mt_openbook_qa_light/arabic_leaderboard_arabic_mt_openbook_qa_light.yaml

+group: arabic_leaderboard_arabic_mt_openbook_qa_light
+task:
+  - arabic_mt_openbook_qa_light
+
+aggregate_metric_list:
+  - metric: acc
+    aggregation: mean
+    weight_by_size: true
+  - metric: acc_norm
+    aggregation: mean
+    weight_by_size: true
+metadata:
+  version: 1.0

lm_eval/tasks/arabic_leaderboard_light/arabic_leaderboard_arabic_mt_openbook_qa_light/arabic_mt_openbook_qa_light.yaml

+task: arabic_mt_openbook_qa_light
+dataset_path: arcee-globe/AlGhafa-Arabic-LLM-Benchmark-Translated-10percent
+dataset_name: openbook_qa_ext_ar
+output_type: multiple_choice
+training_split: null
+validation_split: validation
+test_split: test
+process_docs: !function utils.process_docs
+doc_to_text: "{{query}}"
+doc_to_target: "{{gold}}"
+doc_to_choice: "choices"
+fewshot_split: validation
+fewshot_config:
+  sampler: first_n
+metric_list:
+  - metric: acc
+    aggregation: mean
+    higher_is_better: true
+  - metric: acc_norm
+    aggregation: mean
+    higher_is_better: true
+metadata:
+  version: 1.0

lm_eval/tasks/arabic_leaderboard_light/arabic_leaderboard_arabic_mt_openbook_qa_light/utils.py

+import datasets
+import numpy as np
+
+
+def process_docs(dataset: datasets.Dataset):
+    def _process_doc(doc):
+        question = doc["query"]
+        answer_index = int(doc["label"])
+        # Dynamically determining the choices by excluding '__few_shots', 'query' and 'label'
+        choices_keys = [
+            key for key in doc.keys() if key not in ["query", "label", "__few_shots"]
+        ]
+        choices = [doc[key] for key in choices_keys]
+
+        instruction = "الأسئلة التالية هي أسئلة متعددة الإختيارات مع الجواب الصحيح\n\n"
+        query = f"{instruction}السؤال: {question}\n"
+        for index, choice in enumerate(choices):
+            query += f"{index}) {choice}\n"
+        query += "الإجابة:"
+
+        return {"query": query, "choices": choices, "gold": answer_index}
+
+    return dataset.map(_process_doc)

lm_eval/tasks/arabic_leaderboard_light/arabic_leaderboard_arabic_mt_piqa_light/arabic_leaderboard_arabic_mt_piqa_light.yaml

+group: arabic_leaderboard_arabic_mt_piqa_light
+task:
+  - arabic_mt_piqa_light
+
+aggregate_metric_list:
+  - metric: acc
+    aggregation: mean
+    weight_by_size: true
+  - metric: acc_norm
+    aggregation: mean
+    weight_by_size: true
+metadata:
+  version: 1.0

lm_eval/tasks/arabic_leaderboard_light/arabic_leaderboard_arabic_mt_piqa_light/arabic_mt_piqa_light.yaml

+task: arabic_mt_piqa_light
+dataset_path: arcee-globe/AlGhafa-Arabic-LLM-Benchmark-Translated-10percent
+dataset_name: piqa_ar
+output_type: multiple_choice
+training_split: null
+validation_split: validation
+test_split: test
+process_docs: !function utils.process_docs
+doc_to_text: "{{query}}"
+doc_to_target: "{{gold}}"
+doc_to_choice: "choices"
+fewshot_split: validation
+fewshot_config:
+  sampler: first_n
+metric_list:
+  - metric: acc
+    aggregation: mean
+    higher_is_better: true
+  - metric: acc_norm
+    aggregation: mean
+    higher_is_better: true
+metadata:
+  version: 1.0

lm_eval/tasks/arabic_leaderboard_light/arabic_leaderboard_arabic_mt_piqa_light/utils.py

+import datasets
+import numpy as np
+
+
+def process_docs(dataset: datasets.Dataset):
+    def _process_doc(doc):
+        question = doc["query"]
+        answer_index = int(doc["label"])
+        # Dynamically determining the choices by excluding '__few_shots', 'query' and 'label'
+        choices_keys = [
+            key for key in doc.keys() if key not in ["query", "label", "__few_shots"]
+        ]
+        choices = [doc[key] for key in choices_keys]
+
+        instruction = "الأسئلة التالية هي أسئلة متعددة الإختيارات مع الجواب الصحيح\n\n"
+        query = f"{instruction}السؤال: {question}\n"
+        for index, choice in enumerate(choices):
+            query += f"{index}) {choice}\n"
+        query += "الإجابة:"
+
+        return {"query": query, "choices": choices, "gold": answer_index}
+
+    return dataset.map(_process_doc)

lm_eval/tasks/arabic_leaderboard_light/arabic_leaderboard_arabic_mt_race_light/arabic_leaderboard_arabic_mt_race_light.yaml

+group: arabic_leaderboard_arabic_mt_race_light
+task:
+  - arabic_mt_race_light
+
+aggregate_metric_list:
+  - metric: acc
+    aggregation: mean
+    weight_by_size: true
+  - metric: acc_norm
+    aggregation: mean
+    weight_by_size: true
+metadata:
+  version: 1.0

lm_eval/tasks/arabic_leaderboard_light/arabic_leaderboard_arabic_mt_race_light/arabic_mt_race_light.yaml

+task: arabic_mt_race_light
+dataset_path: arcee-globe/AlGhafa-Arabic-LLM-Benchmark-Translated-10percent
+dataset_name: race_ar
+output_type: multiple_choice
+training_split: null
+validation_split: validation
+test_split: test
+process_docs: !function utils.process_docs
+doc_to_text: "{{query}}"
+doc_to_target: "{{gold}}"
+doc_to_choice: "choices"
+fewshot_split: validation
+fewshot_config:
+  sampler: first_n
+metric_list:
+  - metric: acc
+    aggregation: mean
+    higher_is_better: true
+  - metric: acc_norm
+    aggregation: mean
+    higher_is_better: true
+metadata:
+  version: 1.0

lm_eval/tasks/arabic_leaderboard_light/arabic_leaderboard_arabic_mt_race_light/utils.py

+import datasets
+import numpy as np
+
+
+def process_docs(dataset: datasets.Dataset):
+    def _process_doc(doc):
+        question = doc["query"]
+        answer_index = int(doc["label"])
+        # Dynamically determining the choices by excluding '__few_shots', 'query' and 'label'
+        choices_keys = [
+            key for key in doc.keys() if key not in ["query", "label", "__few_shots"]
+        ]
+        choices = [doc[key] for key in choices_keys]
+
+        instruction = "الأسئلة التالية هي أسئلة متعددة الإختيارات مع الجواب الصحيح\n\n"
+        query = f"{instruction}السؤال: {question}\n"
+        for index, choice in enumerate(choices):
+            query += f"{index}) {choice}\n"
+        query += "الإجابة:"
+
+        return {"query": query, "choices": choices, "gold": answer_index}
+
+    return dataset.map(_process_doc)

lm_eval/tasks/arabic_leaderboard_light/arabic_leaderboard_arabic_mt_sciq_light/arabic_leaderboard_arabic_mt_sciq_light.yaml

+group: arabic_leaderboard_arabic_mt_sciq_light
+task:
+  - arabic_mt_sciq_light
+
+aggregate_metric_list:
+  - metric: acc
+    aggregation: mean
+    weight_by_size: true
+  - metric: acc_norm
+    aggregation: mean
+    weight_by_size: true
+metadata:
+  version: 1.0

lm_eval/tasks/arabic_leaderboard_light/arabic_leaderboard_arabic_mt_sciq_light/arabic_mt_sciq_light.yaml

+task: arabic_mt_sciq_light
+dataset_path: arcee-globe/AlGhafa-Arabic-LLM-Benchmark-Translated-10percent
+dataset_name: sciq_ar
+output_type: multiple_choice
+training_split: null
+validation_split: validation
+test_split: test
+process_docs: !function utils.process_docs
+doc_to_text: "{{query}}"
+doc_to_target: "{{gold}}"
+doc_to_choice: "choices"
+fewshot_split: validation
+fewshot_config:
+  sampler: first_n
+metric_list:
+  - metric: acc
+    aggregation: mean
+    higher_is_better: true
+  - metric: acc_norm
+    aggregation: mean
+    higher_is_better: true
+metadata:
+  version: 1.0

lm_eval/tasks/arabic_leaderboard_light/arabic_leaderboard_arabic_mt_sciq_light/utils.py

+import random
+
+import datasets
+import numpy as np
+
+
+def doc_to_text(doc):
+    instruction = (
+        "بناءً على السياق أدناه، اختر الإجابة الصحيحة للسؤال التالي من قائمة الاقتراحات"
+    )
+    support = doc["support"]
+    question = doc["question"]
+    query = f"""{instruction}
+    السياق:
+    {support}
+    السؤال:
+    {question}
+    الإجابات المحتملة:
+
+    """
+    return query
+
+
+def process_docs(dataset: datasets.Dataset):
+    def _process_doc(doc):
+        correct_answer = doc["correct_answer"]
+        choices = [
+            doc["distractor1"],
+            doc["distractor2"],
+            doc["distractor3"],
+            correct_answer,
+        ]
+
+        # Shuffle the choices
+        random.shuffle(choices)
+
+        answer_index = choices.index(correct_answer)
+
+        return {"query": doc_to_text(doc), "choices": choices, "gold": answer_index}
+
+    return dataset.map(_process_doc)

lm_eval/tasks/arabic_leaderboard_light/arabic_leaderboard_arabic_mt_toxigen_light/arabic_leaderboard_arabic_mt_toxigen_light.yaml

+group: arabic_leaderboard_arabic_mt_toxigen_light
+task:
+  - arabic_mt_toxigen_light
+
+aggregate_metric_list:
+  - metric: acc
+    aggregation: mean
+    weight_by_size: true
+  - metric: acc_norm
+    aggregation: mean
+    weight_by_size: true
+metadata:
+  version: 1.0

lm_eval/tasks/arabic_leaderboard_light/arabic_leaderboard_arabic_mt_toxigen_light/arabic_mt_toxigen_light.yaml

+task: arabic_mt_toxigen_light
+dataset_path: arcee-globe/AlGhafa-Arabic-LLM-Benchmark-Translated-10percent
+dataset_name: toxigen_ar
+output_type: multiple_choice
+training_split: null
+validation_split: validation
+test_split: test
+process_docs: !function utils.process_docs
+doc_to_text: "{{query}}"
+doc_to_target: "{{gold}}"
+doc_to_choice: "choices"
+fewshot_split: validation
+fewshot_config:
+  sampler: first_n
+metric_list:
+  - metric: acc
+    aggregation: mean
+    higher_is_better: true
+  - metric: acc_norm
+    aggregation: mean
+    higher_is_better: true
+metadata:
+  version: 1.0

lm_eval/tasks/arabic_leaderboard_light/arabic_leaderboard_arabic_mt_toxigen_light/utils.py

+import datasets
+import numpy as np
+
+
+def process_docs(dataset: datasets.Dataset) -> datasets.Dataset:
+    def _process_doc(doc):
+        text = doc["text"]
+        label = 1 if ((doc["toxicity_ai"] + doc["toxicity_human"]) > 5.5) else 0
+        instruction = 'هل العبارة التالية تحض على الكراهية؟ قم بالرد بـ"نعم" أو "لا".'
+
+        query = f"""{instruction}
+        العبارة:
+        '{text}'
+        الإجابة:
+        """
+        out_doc = {
+            "query": query,
+            "choices": ["لا", "نعم"],
+            "gold": label,
+        }
+        return out_doc
+
+    return dataset.map(_process_doc)

lm_eval/tasks/arabic_leaderboard_light/arabic_leaderboard_avca_light/arabic_leaderboard_acva_Algeria_light.yaml

+task: arabic_leaderboard_acva_Algeria_light
+dataset_path: arcee-globe/ACVA-10percent
+dataset_name: Algeria
+output_type: multiple_choice
+training_split: null
+validation_split: validation
+test_split: test
+process_docs: !function utils.process_docs
+doc_to_text: "{{query}}"
+doc_to_target: "{{gold}}"
+doc_to_choice: "choices"
+fewshot_split: validation
+fewshot_config:
+  sampler: first_n
+metric_list:
+  - metric: acc
+    aggregation: mean
+    higher_is_better: true
+  - metric: acc_norm
+    aggregation: mean
+    higher_is_better: true
+metadata:
+  version: 1.0

lm_eval/tasks/arabic_leaderboard_light/arabic_leaderboard_avca_light/arabic_leaderboard_acva_Ancient_Egypt_light.yaml

+task: arabic_leaderboard_acva_Ancient_Egypt_light
+dataset_path: arcee-globe/ACVA-10percent
+dataset_name: Ancient_Egypt
+output_type: multiple_choice
+training_split: null
+validation_split: validation
+test_split: test
+process_docs: !function utils.process_docs
+doc_to_text: "{{query}}"
+doc_to_target: "{{gold}}"
+doc_to_choice: "choices"
+fewshot_split: validation
+fewshot_config:
+  sampler: first_n
+metric_list:
+  - metric: acc
+    aggregation: mean
+    higher_is_better: true
+  - metric: acc_norm
+    aggregation: mean
+    higher_is_better: true
+metadata:
+  version: 1.0

lm_eval/tasks/arabic_leaderboard_light/arabic_leaderboard_avca_light/arabic_leaderboard_acva_Arab_Empire_light.yaml

+task: arabic_leaderboard_acva_Arab_Empire_light
+dataset_path: arcee-globe/ACVA-10percent
+dataset_name: Arab_Empire
+output_type: multiple_choice
+training_split: null
+validation_split: validation
+test_split: test
+process_docs: !function utils.process_docs
+doc_to_text: "{{query}}"
+doc_to_target: "{{gold}}"
+doc_to_choice: "choices"
+fewshot_split: validation
+fewshot_config:
+  sampler: first_n
+metric_list:
+  - metric: acc
+    aggregation: mean
+    higher_is_better: true
+  - metric: acc_norm
+    aggregation: mean
+    higher_is_better: true
+metadata:
+  version: 1.0

lm_eval/tasks/arabic_leaderboard_light/arabic_leaderboard_avca_light/arabic_leaderboard_acva_Arabic_Architecture_light.yaml

+task: arabic_leaderboard_acva_Arabic_Architecture_light
+dataset_path: arcee-globe/ACVA-10percent
+dataset_name: Arabic_Architecture
+output_type: multiple_choice
+training_split: null
+validation_split: validation
+test_split: test
+process_docs: !function utils.process_docs
+doc_to_text: "{{query}}"
+doc_to_target: "{{gold}}"
+doc_to_choice: "choices"
+fewshot_split: validation
+fewshot_config:
+  sampler: first_n
+metric_list:
+  - metric: acc
+    aggregation: mean
+    higher_is_better: true
+  - metric: acc_norm
+    aggregation: mean
+    higher_is_better: true
+metadata:
+  version: 1.0