OpenMMReasoner/OpenMMReasoner-ColdStart
OpenMMReasoner-ColdStart is a 7 billion parameter multimodal reasoning model developed by Zhang et al. (EvolvingLMMs-Lab) with a 32768 token context length. It is the supervised fine-tuning (SFT) stage of the OpenMMReasoner project, built upon Qwen2.5-VL-7B-Instruct. This model is specifically designed for multimodal reasoning tasks, achieving state-of-the-art performance on several benchmarks, including MathVista, MathVerse, and WeMath, by emphasizing data quality and a transparent two-stage training recipe.
Loading preview...
OpenMMReasoner-ColdStart: A Foundation for Multimodal Reasoning
OpenMMReasoner-ColdStart is the initial supervised fine-tuning (SFT) stage of the OpenMMReasoner project, a 7 billion parameter multimodal reasoning model developed by Zhang et al. (EvolvingLMMs-Lab). This model is built on the Qwen2.5-VL-7B-Instruct architecture and features a 32768 token context length. It represents a transparent and reproducible approach to multimodal reasoning, focusing on high-quality data curation and a two-stage training strategy.
Key Capabilities
- Robust Multimodal Reasoning: Designed to excel in complex reasoning tasks involving both visual and textual information.
- Strong Foundational Training: Leverages an 874K-sample cold-start dataset with rigorous step-by-step validation, providing a solid base for advanced reasoning.
- Benchmark-Leading Performance: Achieves state-of-the-art results on several challenging multimodal reasoning benchmarks, including 79.5% on MathVista testmini, 63.8% on MathVerse testmini, and 79.0% on WeMath loose, often surpassing larger or more data-intensive models.
- Efficient Data Utilization: Demonstrates high performance with significantly less training data (874K SFT samples) compared to many competing methods, highlighting the importance of data quality and training design.
Good For
- Research in Multimodal AI: Provides a transparent and reproducible recipe for developing and evaluating multimodal reasoning models.
- Applications Requiring Visual-Textual Understanding: Ideal for tasks that demand deep comprehension and logical inference from combined image and text inputs.
- Benchmarking and Comparison: Serves as a strong baseline for evaluating new multimodal reasoning techniques and datasets.
- Developing Specialized Reasoning Agents: Offers a powerful foundation for further fine-tuning or reinforcement learning to create highly specialized multimodal AI systems.