Augmented Grade School Math Problems for Language Model Training

MetaMath_DPO_FewShot is an augmented version of the GSM8K dataset, containing grade school math word problems designed to test mathematical reasoning. The dataset, created by abacusai, partitions problems into queries and step-by-step logical responses. It was last updated on the Hugging Face platform in February 2024.

Use Cases

• Fine-tuning language models for step-by-step mathematical reasoning using the query and response pairs.
• Training models for Direct Preference Optimization (DPO) with the structured question-answer format.
• Benchmarking model performance on diverse grade school math problems from the augmented GSM8K set.
• Augmenting training data for models to improve performance on chain-of-thought reasoning tasks.

Strengths

• Derived from the GSM8K dataset, which contains diverse grade school math problems.
• Data is augmented from the original GSM8K training set, potentially increasing sample variety.

Limitations

• Specific row count, column details, and augmentation methodology are not provided.
• The dataset's size category is listed as '100K<n<1M', indicating a potential but unconfirmed upper bound.

Provenance

Source

abacusai on Hugging Face, based on the GSM8K dataset by Cobbe et al. (2021).

Collection Method

Data augmentation applied to the GSM8K training set.

Freshness

Last updated on the platform in February 2024.