MiroMind-M1 RL Training Performance on AIME24 and AIME25

Training performance data for the MiroMind-M1-RL-7B model on the AIME24 and AIME25 benchmarks. The dataset is associated with a model trained via reinforcement learning with verifiable rewards on 62,000 challenging examples. It is authored by miromind-ai and was last updated in July 2025.

Use Cases

• Analyze the performance of the MiroMind-M1-RL-7B model on the AIME24 benchmark for mathematical reasoning tasks.
• Compare model results between the AIME24 and AIME25 benchmarks to evaluate progress in reasoning capabilities.
• Study the effectiveness of reinforcement learning with verifiable rewards (RLVR) training on 62K examples for problem-solving tasks.

Strengths

• Focuses on a specific, high-quality training subset of 62,000 challenging examples for reinforcement learning.
• Provides performance data on established mathematical reasoning benchmarks (AIME24 and AIME25).
• Dataset is directly linked to a fully open-source model series built on Qwen-2.5.

Limitations

• Dataset contents (e.g., specific columns, sample data, file formats) are not described, limiting usability assessment.
• The raw description is minimal, requiring users to visit an external page for full details.
• The scope is narrowly focused on performance metrics for a single model family on specific benchmarks.

Provenance

Source

miromind-ai on Hugging Face.

Collection Method

Model training performance data, likely generated during the reinforcement learning with verifiable rewards (RLVR) phase.

Freshness

Last updated on 2025-07-22.