peterkchung/commonsense_cot_partial_annotated_prelim

Commonsense QA CoT (Partial, Annotated) - PRELIMINARY

Dataset Summary

This dataset is a human-annotated subset of randomly sampled question-answer entries from the CommonsenseQA dataset (tau/commonsense_qa). The 'rationales' for each QA pair were created using a two-part method. First, Mixtral (mistralai/Mixtral-8x7B-Instruct-v0.1) was used to generate 3 unique CoT (Chain-of-Thought) explanations. Next, human evaluation was applied to distill the random sampling down to a cohesive set of question-answer-rationale triplets.

The working hypothesis, inspired by the research papers listed below, is that a diverse set of CoT rationales passed along with the CommonsenseQA question-answer choices will provide accelerated commonsense reasoning performance on even a relatively small model (<3B parameters).

Additional refinement and annotations to this dataset are to follow.

Background research and inspiration from the following papers:

CommonsenseQA: A Question Answering Challenge Targeting Commonsense Knowledge (https://arxiv.org/abs/1811.00937)
Chain-of-Thought Prompting Elicits Reasoning in Large Language Models (https://arxiv.org/abs/2201.11903)
Specializing Smaller Language Models towards Multi-Step Reasoning (https://arxiv.org/abs/2301.12726)
Orca 2: Teaching Small Language Models How to Reason (https://arxiv.org/abs/2311.11045)
Large Language Models Are Reasoning Teachers (https://arxiv.org/abs/2212.10071)
Teaching Small Language Models to Reason (https://arxiv.org/abs/2212.08410)

Dataset Structure

Languages

The dataset is in English (en).

Data Fields

• id (str): Unique ID.
• question: a string feature.
• question_concept (str): ConceptNet concept associated to the question.
• choices: a dictionary feature containing:
  • label: a string feature.
  • text: a string feature.
• answerKey: a string feature.
• rationale: a string feature.

Data Example

{'id': '1fe48d12b6f6e4e38f4445f3ec60d5c5',
 'question': 'What can happen  to someone too sure of their learning?',
 'question_concept': 'learning',
 'choices': {'label': ['A', 'B', 'C', 'D', 'E'],
              'text': ['growth',
                       'gaining knowledge',
                       'enlightenment',
                       'knowing more',
                       'overconfidence']},
  'answerKey': 'E',
  'rationale':  'When someone is too sure of their learning, they become '
                'overconfident, thinking that they know everything. This can '
                'prevent them from learning more, as they stop seeking new '
                'knowledge and ideas. They might also miss out on '
                'enlightenment, as they close themselves off to new '
                'perspectives. Overall, their growth might be stunted, as they '
                'stop challenging themselves and expanding their '
                'understanding. So, out of the given choices, the most '
                'appropriate answer is overconfidence.'}

Source Data

• Data: https://huggingface.co/datasets/tau/commonsense_qa
• Homepage: https://www.tau-nlp.org/commonsenseqa
• Repository: https://github.com/jonathanherzig/commonsenseqa
• Paper: https://arxiv.org/abs/1811.00937

Licensing Information

The dataset is licensed under the MIT License.