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_complete/arabic_leaderboard_arabic_mt_arc_easy/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_complete/arabic_leaderboard_arabic_mt_boolq/arabic_leaderboard_arabic_mt_boolq.yaml

+group: arabic_leaderboard_arabic_mt_boolq
+task:
+  - arabic_mt_boolq
+
+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_complete/arabic_leaderboard_arabic_mt_boolq/arabic_mt_boolq.yaml

+task: arabic_mt_boolq
+dataset_path: OALL/AlGhafa-Arabic-LLM-Benchmark-Translated
+dataset_name: boolq_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_complete/arabic_leaderboard_arabic_mt_boolq/utils.py

+import datasets
+import numpy as np
+
+
+def process_docs(dataset: datasets.Dataset):
+    def _process_doc(doc):
+        question = doc["question"]
+        passage = doc["passage"]
+        instruction = "بناء على المقطع التالي، أجب عن السؤال ب نعم أو لا"
+        query = f"""{instruction}
+        المقطع :
+        {passage}
+        السؤال:
+        {question}
+        الإجابة:
+        """
+
+        return {
+            "query": query,
+            "choices": ["نعم", "لا"],
+            "gold": 0 if doc["answer"] else 1,
+        }
+
+    return dataset.map(_process_doc)

lm_eval/tasks/arabic_leaderboard_complete/arabic_leaderboard_arabic_mt_copa/arabic_leaderboard_arabic_mt_copa.yaml

+group: arabic_leaderboard_arabic_mt_copa
+task:
+  - arabic_mt_copa
+
+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_complete/arabic_leaderboard_arabic_mt_copa/arabic_mt_copa.yaml

+task: arabic_mt_copa
+dataset_path: OALL/AlGhafa-Arabic-LLM-Benchmark-Translated
+dataset_name: copa_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_complete/arabic_leaderboard_arabic_mt_copa/utils.py

+import datasets
+import numpy as np
+
+
+def process_docs(dataset: datasets.Dataset):
+    def _process_doc(doc):
+        premise = doc["premise"]
+        choices = [doc["choice1"], doc["choice2"]]
+        question_map = {"cause": "لأن", "effect": "لذلك"}
+        question = question_map[doc["question"]]
+        answer = doc["label"]
+
+        query = "{}، {} :\n0) {}\n1) {}\nالإجابة:".format(
+            premise, question, choices[0], choices[1]
+        )
+
+        return {"query": query, "choices": choices, "gold": answer}
+
+    return dataset.map(_process_doc)

lm_eval/tasks/arabic_leaderboard_complete/arabic_leaderboard_arabic_mt_hellaswag/arabic_leaderboard_arabic_mt_hellaswag.yaml

+group: arabic_leaderboard_arabic_mt_hellaswag
+task:
+  - arabic_mt_hellaswag
+
+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_complete/arabic_leaderboard_arabic_mt_hellaswag/arabic_mt_hellaswag.yaml

+task: arabic_mt_hellaswag
+dataset_path: OALL/AlGhafa-Arabic-LLM-Benchmark-Translated
+dataset_name: hellaswag_okapi_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_complete/arabic_leaderboard_arabic_mt_hellaswag/utils.py

+import re
+
+import datasets
+import numpy as np
+
+
+def process_docs(dataset: datasets.Dataset):
+    def _process_doc(doc):
+        ctx = re.sub(r"\[.*?\]", "", doc["ctx"])  # Remove latin words within brackets
+        endings = [
+            re.sub(r"\[.*?\]", "", e) for e in eval(doc["endings"])
+        ]  # endings is a string representation of a list
+        answer_index = doc["label"]
+        instruction = (
+            "بناء على السياق التالي، اختر النهاية الصحيحة من الاقتراحات التالية"
+        )
+
+        query = f"""{instruction}
+        السياق:
+        {ctx}
+        الاقتراحات:
+
+        """
+        for i, ending in enumerate(endings):
+            query += f"{i}) {ending}\n"
+        query += "الإجابة:"
+
+        return {"query": query, "choices": endings, "gold": answer_index}
+
+    return dataset.map(_process_doc)

lm_eval/tasks/arabic_leaderboard_complete/arabic_leaderboard_arabic_mt_mmlu/arabic_leaderboard_arabic_mt_mmlu.yaml

+group: arabic_leaderboard_arabic_mt_mmlu
+task:
+  - arabic_mt_mmlu
+
+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_complete/arabic_leaderboard_arabic_mt_mmlu/arabic_mt_mmlu.yaml

+task: arabic_mt_mmlu
+dataset_path: OALL/AlGhafa-Arabic-LLM-Benchmark-Translated
+dataset_name: mmlu_okapi_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_complete/arabic_leaderboard_arabic_mt_mmlu/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_complete/arabic_leaderboard_arabic_mt_openbook_qa/arabic_leaderboard_arabic_mt_openbook_qa.yaml

+group: arabic_leaderboard_arabic_mt_openbook_qa
+task:
+  - arabic_mt_openbook_qa
+
+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_complete/arabic_leaderboard_arabic_mt_openbook_qa/arabic_mt_openbook_qa.yaml

+task: arabic_mt_openbook_qa
+dataset_path: OALL/AlGhafa-Arabic-LLM-Benchmark-Translated
+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_complete/arabic_leaderboard_arabic_mt_openbook_qa/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_complete/arabic_leaderboard_arabic_mt_piqa/arabic_leaderboard_arabic_mt_piqa.yaml

+group: arabic_leaderboard_arabic_mt_piqa
+task:
+  - arabic_mt_piqa
+
+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_complete/arabic_leaderboard_arabic_mt_piqa/arabic_mt_piqa.yaml

+task: arabic_mt_piqa
+dataset_path: OALL/AlGhafa-Arabic-LLM-Benchmark-Translated
+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_complete/arabic_leaderboard_arabic_mt_piqa/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_complete/arabic_leaderboard_arabic_mt_race/arabic_leaderboard_arabic_mt_race.yaml

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