Merge branch 'main' into prefill

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

lm_eval/tasks/arabic_leaderboard_complete/arabic_leaderboard_arabic_mt_race/arabic_mt_race.yaml

+task: arabic_mt_race
+dataset_path: OALL/AlGhafa-Arabic-LLM-Benchmark-Translated
+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_complete/arabic_leaderboard_arabic_mt_race/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_sciq/arabic_leaderboard_arabic_mt_sciq.yaml

+group: arabic_leaderboard_arabic_mt_sciq
+task:
+  - arabic_mt_sciq
+
+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_sciq/arabic_mt_sciq.yaml

+task: arabic_mt_sciq
+dataset_path: OALL/AlGhafa-Arabic-LLM-Benchmark-Translated
+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_complete/arabic_leaderboard_arabic_mt_sciq/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_complete/arabic_leaderboard_arabic_mt_toxigen/arabic_leaderboard_arabic_mt_toxigen.yaml

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