pszemraj/unified-mcqa

Unified MCQA: Aggregated Multiple-Choice Datasets

A standardized collection of multiple-choice question answering datasets organized by the number of answer choices (typically 2, 3, 4, 5, or 8). Only 'train' splits (or equivalent, like MMLU's 'auxiliary_train') were included to avoid contamination with evaluation sets.

The intended use case is to include phase(s) of training on MCQ for encoders as part of a 'multi-task pretraining' with an adapted version of the multiple choice example script.

Dataset Structure

Data Fields

• context (string): Supporting text for the question (may be empty).
• question (string): The multiple-choice question.
• choices (List[string]): Potential answer options, with prefixes like "a)" removed.
• label (int): 0-indexed integer representing the correct answer choice.
• source_dataset (string): Identifier for the original source dataset.

Configurations

The dataset is organized into configurations based on the number of choices per question after processing:

• 2-choice: Examples with 2 answer choices (e.g., from PIQA, Winogrande, COPA).
• 3-choice: Examples with 3 answer choices (e.g., from Social IQa).
• 4-choice: Examples with 4 answer choices (e.g., from RACE, ARC, OpenBookQA, SWAG, Cosmos QA, DREAM, Quail, MedMCQA, MMLU, QuALITY, MCTest).
• 5-choice: Examples with 5 answer choices (e.g., from CommonsenseQA, MathQA).
• 8-choice: Examples with 8 answer choices (e.g., from QASC).

Only train splits are currently available for each configuration.

Quick Usage

from datasets import load_dataset

# Load 4-choice examples
# Replace 'pszemraj/unified-mcqa' with the actual Hub path if different
ds = load_dataset("pszemraj/unified-mcqa", name="4-choice", split="train")

# Example structure
print(ds)
# Example Output:
# {
#   'context': 'Article text here...',
#   'question': 'What is the main idea?',
#   'choices': ['Choice text A', 'Choice text B', 'Choice text C', 'Choice text D'],
#   'label': 2,
#   'source_dataset': 'race'
# }

# Load 5-choice examples
ds_5 = load_dataset("pszemraj/unified-mcqa", name="5-choice", split="train")
print(ds_5)
# Example Output (MathQA):
# {
#  'context': '',
#  'question': 'a monkey ascends a greased pole 26 meters high. he ascends 2 meters in the first minute and then slips down 1 meter in the alternate minute. if this pattern continues until he climbs the pole, in how many minutes would he reach at the top of the pole?',
#  'choices': ['50 th minute', '41 st minute', '45 th minute', '42 nd minute', '49 th minute'],
#  'label': 4,
#  'source_dataset': 'math_qa'
# }

Source Datasets

This dataset aggregates examples from the following Hugging Face datasets (using specified paths/configs):

• commonsense_qa (path: commonsense_qa)
• sciq (path: sciq)
• math_qa (path: math_qa)
• swag (path: swag, name: regular)
• hellaswag (path: hellaswag)
• social_i_qa (path: social_i_qa)
• cosmos_qa (path: cosmos_qa)
• piqa (path: piqa)
• winogrande (path: winogrande, name: winogrande_xl)
• dream (path: dream)
• quail (path: quail)
• ai2_arc (name: ARC-Challenge) -> arc_challenge
• ai2_arc (name: ARC-Easy) -> arc_easy
• openbookqa (path: openbookqa, name: main)
• cais/mmlu (name: all, split: auxiliary_train) -> mmlu
• lighteval/QuALITY (path: lighteval/QuALITY) -> quality
• qasc (path: qasc)
• codah (path: codah, name: codah)
• super_glue (name: copa) -> copa
• copenlu/mctest_corrected (name: mc500) -> mctest_corrected
• openlifescienceai/medmcqa (path: openlifescienceai/medmcqa) -> medmcqa
• race (path: race, name: all)

Creation Process

The aggregation script (create_unified_mcqa.py) performs the following steps for each source dataset:

1. Loads the specified split (usually train).
2. Applies a custom canonicalization function (SPECIAL) or a generic mapping (COLMAP) to standardize field names (context, question, choices, label) and structures.
   • For math_qa, the Problem field is mapped to question, and the Rationale is ignored. Choice prefixes (e.g., "a )") are removed.
3. Converts all labels to 0-indexed integers.
4. Adds a source_dataset field to identify the origin.
5. Filters out malformed or skipped examples.
6. Casts the resulting examples to the TARGET_FEATURES schema.
7. Groups the processed datasets by the number of choices found in each example.
8. Concatenates datasets within each group.

License

This aggregated dataset combines data from multiple sources, each with its own license. The licenses are generally permissive (MIT, Apache 2.0, CC BY variants, ODC-By 1.0, Custom Permissive).

Users should consult the original licenses for each source dataset included in their specific use case, especially before commercial use. The aggregation script attempts to log the license found for each dataset during processing.

Citation

If using this dataset, please cite the original datasets that contribute to your specific use case. Refer to the Hugging Face Hub pages for the source datasets listed above for their respective citation information.