GSM8K Math Problem Dataset for ES and GRPO Comparison

AlphaXiv's dataset contains prepared splits of the GSM8K grade school math problems for comparing Evolution Strategies and Group Relative Policy Optimization methods in LLM fine-tuning. It includes 6,725 training samples, 1,867 validation samples, and 200 test samples. The dataset was last updated on March 5, 2026.

Use Cases

• Fine-tuning LLMs for mathematical reasoning using the 'data' input field containing math problems.
• Benchmarking Evolution Strategies against Group Relative Policy Optimization using the provided training, validation, and test splits.
• Evaluating model performance on grade school math problems via the reserved test set of 200 samples.

Strengths

• Contains 6,725 training examples for model fine-tuning.
• Provides a structured split with 1,867 validation and 200 test samples for evaluation.

Limitations

• Dataset size is moderate, with under 10,000 total samples.
• Specific column structure and data format details are not fully described.

Provenance

Source

AlphaXiv on Hugging Face.

Collection Method

Prepared splits of the GSM8K dataset.

Freshness

Last updated on March 5, 2026.