GM-PRM-20K: Multimodal Math Problem Solutions for Training Reward Models

GM-PRM-20K is a training dataset for a Generative Multimodal Process Reward Model, as described in arXiv:2508.04088. It contains full multi-step solutions to multimodal math problems, used to train the model zijinghuafen/GM-PRM. The dataset was accepted at the 4th Workshop on Advances in Language and Vision Research (ALVR), in conjunction with ACL 2026.

Use Cases

• Training multimodal reward models based on step-by-step solution processes.
• Fine-tuning large language models for mathematical reasoning tasks using paired problem-solution data.
• Benchmarking the performance of generative AI on complex, multimodal math problems.
• Researching the alignment of AI reasoning processes with human-like problem-solving steps.

Strengths

• Dataset is directly linked to a published research paper (arXiv:2508.04088) and an accepted workshop presentation.
• It is the exact dataset used to train a specific, publicly released model (zijinghuafen/GM-PRM).

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count is unknown, which may limit suitability assessment.
• Description metadata is limited; actual data quality and structure require manual inspection after download.

Provenance

Source

Author zijinghuafen on Hugging Face.

Collection Method

Created as a training dataset for the GM-PRM model, likely derived from multimodal math problems.

Freshness

Last updated 2026-06-10 04:30:29; freshness should be verified.