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_avca/arabic_leaderboard_acva_Tunisia.yaml

+task: arabic_leaderboard_acva_Tunisia
+dataset_path: OALL/ACVA
+dataset_name: Tunisia
+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/arabic_leaderboard_acva_United_Arab_Emirates.yaml

+task: arabic_leaderboard_acva_United_Arab_Emirates
+dataset_path: OALL/ACVA
+dataset_name: United_Arab_Emirates
+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/arabic_leaderboard_acva_Yemen.yaml

+task: arabic_leaderboard_acva_Yemen
+dataset_path: OALL/ACVA
+dataset_name: Yemen
+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/arabic_leaderboard_acva_communication.yaml

+task: arabic_leaderboard_acva_communication
+dataset_path: OALL/ACVA
+dataset_name: communication
+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/arabic_leaderboard_acva_computer_and_phone.yaml

+task: arabic_leaderboard_acva_computer_and_phone
+dataset_path: OALL/ACVA
+dataset_name: computer_and_phone
+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/arabic_leaderboard_acva_daily_life.yaml

+task: arabic_leaderboard_acva_daily_life
+dataset_path: OALL/ACVA
+dataset_name: daily_life
+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/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