Math Question Responses for Reward Model Training

356,058 unique math questions, each paired with 6 different responses, resulting in 2,136,348 total examples. This data was used by NVIDIA to train the AceMath-7/72B-RM reward model for mathematical outcome evaluation. The dataset was released in January 2025.

Use Cases

• Train a reward model to score the quality of different 'response' texts for a given 'math question'.
• Analyze response diversity by comparing the 6 different answers generated per unique question.
• Benchmark the performance of math-specific reward models against proprietary models using the provided evaluation results.
• Fine-tune a language model for mathematical instruction following using the question-response pairs.

Strengths

• Contains 2,136,348 total training examples.
• Features 356,058 unique math problems, providing broad coverage.
• Each question has 6 associated responses, enabling comparative analysis.

Limitations

• Specific column names and data schema are not provided in the input.
• The source and difficulty distribution of the 356,058 math questions is unknown.
• No information on potential biases in question topics or response generation methods.

Provenance

Source

NVIDIA, as part of the AceMath project.

Collection Method

Created for training the AceMath-7/72B-RM reward model; specific collection/generation method not detailed.

Freshness

Last updated on the platform in January 2025.