Mathematical Reasoning Data For Flawed-Aware Policy Optimization

A dataset for training the FAPO-32B model in flawed-aware policy optimization for reasoning tasks. The training data originates from DAPO-Math-17K, duplicated 20 times, while test data mixes duplicated AIME24, AIME25, and GPQA-Diamond problems. It was created by user 'dyyyyyyyy' and last updated in October 2025.

Use Cases

• Training flawed-aware policy optimization models using the '\boxed{}' instruction format from the training data.
• Benchmarking model performance on the AIME24 and AIME25 competition problems in the test set.
• Evaluating reasoning reliability on the GPQA-Diamond subset within the test data.
• Analyzing the effect of data duplication strategies (20x, 32x, 4x) on model training and evaluation.

Strengths

• Training data is explicitly sourced from the established DAPO-Math-17K dataset.
• Test set combines multiple challenging benchmarks: AIME24, AIME25, and GPQA-Diamond.
• Dataset is actively maintained, with a last update in October 2025.

Limitations

• Exact row counts, column names, and dataset size are unknown.
• The dataset is highly specialized for a specific training method (FAPO), limiting general applicability.
• Data is heavily duplicated, which may introduce biases or reduce diversity.

Provenance

Source

Hugging Face user dyyyyyyyy.

Collection Method

Compiled from DAPO-Math-17K and test mixtures from AIME and GPQA benchmarks, with applied duplication.

Time Range

null

Freshness

Last updated 2025-10-28.

Geography

null