Merge branch 'main' into prefill

lm_eval/tasks/arabic_leaderboard_complete/arabic_leaderboard_avca/arabic_leaderboard_acva_entertainment.yaml

+task: arabic_leaderboard_acva_entertainment
+dataset_path: OALL/ACVA
+dataset_name: entertainment
+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_avca/utils.py

+import datasets
+import numpy as np
+
+
+def process_docs(dataset: datasets.Dataset):
+    def _process_doc(doc):
+        question = doc["question"]
+        answer = doc["answer"]
+
+        return {
+            "query": f"السؤال: {question}\nالإجابة:",
+            "choices": ["صح", "خطأ"],
+            "gold": ["صح", "خطأ"].index(answer),
+        }
+
+    return dataset.map(_process_doc)

lm_eval/tasks/arabic_leaderboard_complete/arabic_leaderboard_complete.yaml

+group: arabic_leaderboard_complete
+task:
+  - arabic_leaderboard_acva
+  - arabic_leaderboard_alghafa
+  - arabic_leaderboard_arabic_exams
+  - arabic_leaderboard_arabic_mt_arc_challenge
+  - arabic_leaderboard_arabic_mt_arc_easy
+  - arabic_leaderboard_arabic_mt_boolq
+  - arabic_leaderboard_arabic_mt_hellaswag
+  - arabic_leaderboard_arabic_mt_mmlu
+  - arabic_leaderboard_arabic_mt_copa
+  - arabic_leaderboard_arabic_mt_openbook_qa
+  - arabic_leaderboard_arabic_mt_piqa
+  - arabic_leaderboard_arabic_mt_race
+  - arabic_leaderboard_arabic_mt_sciq
+  - arabic_leaderboard_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

lm_eval/tasks/arabic_leaderboard_light/README.md

+# Arabic Leaderboard Light
+
+Title: Open Arabic LLM Leaderboard Light
+
+This leaderboard follows all the details as in [`arabic_leaderboard_complete`](../arabic_leaderboard_complete), except that a light version - 10% random sample of the test set of each benchmark - is used to test the language models.
+
+NOTE: In ACVA benchmark, there is Yemen subset, and it is a small dataset - it has only 10 samples in the test split. So, for this specific subset dataset, to have more reliable results, we consider the original dataset, instead of 10% of its test samples.
+
+### Checklist
+
+For adding novel benchmarks/datasets to the library:
+* [ ] Is the task an existing benchmark in the literature?
+  * [ ] Have you referenced the original paper that introduced the task?
+  * [ ] If yes, does the original paper provide a reference implementation? If so, have you checked against the reference implementation and documented how to run such a test?
+
+
+If other tasks on this dataset are already supported:
+* [ ] Is the "Main" variant of this task clearly denoted?
+* [ ] Have you provided a short sentence in a README on what each new variant adds / evaluates?
+* [ ] Have you noted which, if any, published evaluation setups are matched by this variant?

lm_eval/tasks/arabic_leaderboard_light/arabic_leaderboard_alghafa_light/arabic_leaderboard_alghafa_light.yaml

+group: arabic_leaderboard_alghafa_light
+task:
+  - arabic_leaderboard_alghafa_mcq_exams_test_ar_light
+  - arabic_leaderboard_alghafa_meta_ar_dialects_light
+  - arabic_leaderboard_alghafa_meta_ar_msa_light
+  - arabic_leaderboard_alghafa_multiple_choice_facts_truefalse_balanced_task_light
+  - arabic_leaderboard_alghafa_multiple_choice_grounded_statement_soqal_task_light
+  - arabic_leaderboard_alghafa_multiple_choice_grounded_statement_xglue_mlqa_task_light
+  - arabic_leaderboard_alghafa_multiple_choice_rating_sentiment_no_neutral_task_light
+  - arabic_leaderboard_alghafa_multiple_choice_rating_sentiment_task_light
+  - arabic_leaderboard_alghafa_multiple_choice_sentiment_task_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_alghafa_light/arabic_leaderboard_alghafa_mcq_exams_test_ar_light.yaml

+task: arabic_leaderboard_alghafa_mcq_exams_test_ar_light
+dataset_path: arcee-globe/AlGhafa-Arabic-LLM-Benchmark-Native-10percent
+dataset_name: mcq_exams_test_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_alghafa_light/arabic_leaderboard_alghafa_meta_ar_dialects_light.yaml

+task: arabic_leaderboard_alghafa_meta_ar_dialects_light
+dataset_path: arcee-globe/AlGhafa-Arabic-LLM-Benchmark-Native-10percent
+dataset_name: meta_ar_dialects
+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_alghafa_light/arabic_leaderboard_alghafa_meta_ar_msa_light.yaml

+task: arabic_leaderboard_alghafa_meta_ar_msa_light
+dataset_path: arcee-globe/AlGhafa-Arabic-LLM-Benchmark-Native-10percent
+dataset_name: meta_ar_msa
+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_alghafa_light/arabic_leaderboard_alghafa_multiple_choice_facts_truefalse_balanced_task_light.yaml

+task: arabic_leaderboard_alghafa_multiple_choice_facts_truefalse_balanced_task_light
+dataset_path: arcee-globe/AlGhafa-Arabic-LLM-Benchmark-Native-10percent
+dataset_name: multiple_choice_facts_truefalse_balanced_task
+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_alghafa_light/arabic_leaderboard_alghafa_multiple_choice_grounded_statement_soqal_task_light.yaml

+task: arabic_leaderboard_alghafa_multiple_choice_grounded_statement_soqal_task_light
+dataset_path: arcee-globe/AlGhafa-Arabic-LLM-Benchmark-Native-10percent
+dataset_name: multiple_choice_grounded_statement_soqal_task
+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_alghafa_light/arabic_leaderboard_alghafa_multiple_choice_grounded_statement_xglue_mlqa_task_light.yaml

+task: arabic_leaderboard_alghafa_multiple_choice_grounded_statement_xglue_mlqa_task_light
+dataset_path: arcee-globe/AlGhafa-Arabic-LLM-Benchmark-Native-10percent
+dataset_name: multiple_choice_grounded_statement_xglue_mlqa_task
+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_alghafa_light/arabic_leaderboard_alghafa_multiple_choice_rating_sentiment_no_neutral_task_light.yaml

+task: arabic_leaderboard_alghafa_multiple_choice_rating_sentiment_no_neutral_task_light
+dataset_path: arcee-globe/AlGhafa-Arabic-LLM-Benchmark-Native-10percent
+dataset_name: multiple_choice_rating_sentiment_no_neutral_task
+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_alghafa_light/arabic_leaderboard_alghafa_multiple_choice_rating_sentiment_task_light.yaml

+task: arabic_leaderboard_alghafa_multiple_choice_rating_sentiment_task_light
+dataset_path: arcee-globe/AlGhafa-Arabic-LLM-Benchmark-Native-10percent
+dataset_name: multiple_choice_rating_sentiment_task
+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_alghafa_light/arabic_leaderboard_alghafa_multiple_choice_sentiment_task_light.yaml

+task: arabic_leaderboard_alghafa_multiple_choice_sentiment_task_light
+dataset_path: arcee-globe/AlGhafa-Arabic-LLM-Benchmark-Native-10percent
+dataset_name: multiple_choice_sentiment_task
+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_alghafa_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_exams_light/arabic_exams_light.yaml

+task: arabic_exams_light
+dataset_path: arcee-globe/Arabic_EXAMS-10percent
+dataset_name: default
+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_exams_light/arabic_leaderboard_arabic_exams_light.yaml

+group: arabic_leaderboard_arabic_exams_light
+task:
+  - arabic_exams_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_exams_light/utils.py

+import datasets
+import numpy as np
+
+
+# fmt: off
+LETTER_INDICES_AR = ["أ", "ب", "ج", "د", "هـ", "و", "ز", "ح", "ط", "ي", "ك", "ل", "م", "ن", "س", "ع", "ف", "ص", "ق", "ر", "ش", "ت", "ث", "خ", "ذ", "ض", "ظ", "غ"]
+# fmt: on
+
+
+# fmt: off
+LETTER_INDICES = ["A", "B", "C", "D", "E", "F", "G", "H", "I", "J", "K", "L", "M", "N", "O", "P", "Q", "R", "S", "T", "U", "V", "W", "X", "Y", "Z"]
+# fmt: on
+
+
+def process_docs(dataset: datasets.Dataset):
+    def _process_doc(doc):
+        topic = doc["subject"]
+        question = doc["question"]
+        choices = [doc["A"], doc["B"], doc["C"], doc["D"]]
+        choices_formatted = [
+            f" {LETTER_INDICES_AR[i]}) {choice}\n" for i, choice in enumerate(choices)
+        ]
+        answer = doc["answer"]
+        answer_index = LETTER_INDICES.index(answer)
+
+        instruction = f"الأسئلة التالية هي أسئلة متعددة الإختيارات مع الجواب الصحيح حول {topic.replace('_', ' ')}. \n\n"
+        query = f"{instruction}السؤال: {question}\n"
+        query += "\n".join(choices_formatted)
+        query += "\nالإجابة:"
+
+        return {"query": query, "choices": LETTER_INDICES_AR[:4], "gold": answer_index}
+
+    return dataset.map(_process_doc)

lm_eval/tasks/arabic_leaderboard_light/arabic_leaderboard_arabic_mmlu_light/arabic_leaderboard_arabic_mmlu_abstract_algebra_light.yaml

+task: arabic_leaderboard_arabic_mmlu_abstract_algebra_light
+dataset_path: arcee-globe/Arabic_MMLU-10percent
+dataset_name: abstract_algebra
+output_type: multiple_choice
+training_split: null
+validation_split: dev
+test_split: test
+process_docs: !function utils.process_docs
+doc_to_text: "{{query}}"
+doc_to_target: "{{gold}}"
+doc_to_choice: "choices"
+fewshot_split: dev
+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_mmlu_light/arabic_leaderboard_arabic_mmlu_anatomy_light.yaml

+task: arabic_leaderboard_arabic_mmlu_anatomy_light
+dataset_path: arcee-globe/Arabic_MMLU-10percent
+dataset_name: anatomy
+output_type: multiple_choice
+training_split: null
+validation_split: dev
+test_split: test
+process_docs: !function utils.process_docs
+doc_to_text: "{{query}}"
+doc_to_target: "{{gold}}"
+doc_to_choice: "choices"
+fewshot_split: dev
+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