Merge pull request #852 from EleutherAI/bigbench

[Refactor] bigbench

lm_eval/tasks/bigbench/multiple_choice/symbol_interpretation.yaml

+# Generated by utils.py
+dataset_name: symbol_interpretation_zero_shot
+include: ../multiple_choice_template_yaml
+task: bigbench_symbol_interpretation_multiple_choice

lm_eval/tasks/bigbench/multiple_choice/temporal_sequences.yaml

+# Generated by utils.py
+dataset_name: temporal_sequences_zero_shot
+include: ../multiple_choice_template_yaml
+task: bigbench_temporal_sequences_multiple_choice

lm_eval/tasks/bigbench/multiple_choice/tense.yaml

+# Generated by utils.py
+dataset_name: tense_zero_shot
+include: ../multiple_choice_template_yaml
+task: bigbench_tense_multiple_choice

lm_eval/tasks/bigbench/multiple_choice/timedial.yaml

+# Generated by utils.py
+dataset_name: timedial_zero_shot
+include: ../multiple_choice_template_yaml
+task: bigbench_timedial_multiple_choice

lm_eval/tasks/bigbench/multiple_choice/topical_chat.yaml

+# Generated by utils.py
+dataset_name: topical_chat_zero_shot
+include: ../multiple_choice_template_yaml
+task: bigbench_topical_chat_multiple_choice

lm_eval/tasks/bigbench/multiple_choice/tracking_shuffled_objects.yaml

+# Generated by utils.py
+dataset_name: tracking_shuffled_objects_zero_shot
+include: ../multiple_choice_template_yaml
+task: bigbench_tracking_shuffled_objects_multiple_choice

lm_eval/tasks/bigbench/multiple_choice/understanding_fables.yaml

+# Generated by utils.py
+dataset_name: understanding_fables_zero_shot
+include: ../multiple_choice_template_yaml
+task: bigbench_understanding_fables_multiple_choice

lm_eval/tasks/bigbench/multiple_choice/undo_permutation.yaml

+# Generated by utils.py
+dataset_name: undo_permutation_zero_shot
+include: ../multiple_choice_template_yaml
+task: bigbench_undo_permutation_multiple_choice

lm_eval/tasks/bigbench/multiple_choice/unit_conversion.yaml

+# Generated by utils.py
+dataset_name: unit_conversion_zero_shot
+include: ../multiple_choice_template_yaml
+task: bigbench_unit_conversion_multiple_choice

lm_eval/tasks/bigbench/multiple_choice/unit_interpretation.yaml

+# Generated by utils.py
+dataset_name: unit_interpretation_zero_shot
+include: ../multiple_choice_template_yaml
+task: bigbench_unit_interpretation_multiple_choice

lm_eval/tasks/bigbench/multiple_choice/unnatural_in_context_learning.yaml

+# Generated by utils.py
+dataset_name: unnatural_in_context_learning_zero_shot
+include: ../multiple_choice_template_yaml
+task: bigbench_unnatural_in_context_learning_multiple_choice

lm_eval/tasks/bigbench/multiple_choice/vitaminc_fact_verification.yaml

+# Generated by utils.py
+dataset_name: vitaminc_fact_verification_zero_shot
+include: ../multiple_choice_template_yaml
+task: bigbench_vitaminc_fact_verification_multiple_choice

lm_eval/tasks/bigbench/multiple_choice/what_is_the_tao.yaml

+# Generated by utils.py
+dataset_name: what_is_the_tao_zero_shot
+include: ../multiple_choice_template_yaml
+task: bigbench_what_is_the_tao_multiple_choice

lm_eval/tasks/bigbench/multiple_choice/which_wiki_edit.yaml

+# Generated by utils.py
+dataset_name: which_wiki_edit_zero_shot
+include: ../multiple_choice_template_yaml
+task: bigbench_which_wiki_edit_multiple_choice

lm_eval/tasks/bigbench/multiple_choice/winowhy.yaml

+# Generated by utils.py
+dataset_name: winowhy_zero_shot
+include: ../multiple_choice_template_yaml
+task: bigbench_winowhy_multiple_choice

lm_eval/tasks/bigbench/multiple_choice/word_sorting.yaml

+# Generated by utils.py
+dataset_name: word_sorting_zero_shot
+include: ../multiple_choice_template_yaml
+task: bigbench_word_sorting_multiple_choice

lm_eval/tasks/bigbench/multiple_choice/word_unscrambling.yaml

+# Generated by utils.py
+dataset_name: word_unscrambling_zero_shot
+include: ../multiple_choice_template_yaml
+task: bigbench_word_unscrambling_multiple_choice

lm_eval/tasks/bigbench/multiple_choice_template_yaml

+group: bigbench
+dataset_path: bigbench # will switch to `hails/bigbench` when all tasks are pushed
+dataset_kwargs:
+  # num_shots: 0 # TODO: num of shots for `bigbench` HF dataset should be controlled through this, not through the typical methods
+  # subtask_name: null
+output_type: multiple_choice
+test_split: default
+doc_to_text: inputs
+doc_to_target: "{{multiple_choice_targets.index(targets[0])}}"
+doc_to_choice: "{{multiple_choice_targets}}"
+metric_list:
+  - metric: acc
+  # TODO: brier score and other metrics

lm_eval/tasks/bigbench/push_bigbench_dataset.py

+"""
+A utility script that pushes all Bigbench subtasks from their form in the `bigbench` HF dataset
+into `{org name}/bigbench`.
+
+Prior to running, log into HF Hub for the target HF hub org via `huggingface-cli login`.
+
+Requires the installation of
+`pip install "bigbench @ https://storage.googleapis.com/public_research_data/bigbench/bigbench-0.0.1.tar.gz"`
+and is included so that the bigbench dependency can be avoided.
+"""
+from tqdm import tqdm
+import datasets
+
+import bigbench.api.util as bb_utils
+
+
+all_task_names = bb_utils.get_all_json_task_names()
+
+num_shots = [0]
+
+for shots in num_shots:
+    for task_name in tqdm(all_task_names):
+        try:
+            print(f"Loading '{task_name}' with num_shots={shots}...")
+            task_ds = datasets.load_dataset("bigbench", name=task_name, num_shots=shots)
+
+            print(f"Pushing '{task_name}' with num_shots={shots}...")
+            task_ds.push_to_hub("hails/bigbench", task_name + "_zero_shot")
+
+            del task_ds
+        except Exception as e:
+            raise e