Awaker2.5-R1 is a multimodal large model developed by Metabrain AGI, featuring complex reasoning and long-chain thinking capabilities implemented through GRPO.
Purpose This model is aimed at tasks requiring multimodal understanding and mathematical reasoning. The information shows that it enhances mathematical reasoning ability through training and achieves performance on MathVista close to the OpenAI o1 model.
Main Capabilities
- Uses Qwen2.5-VL-7B as the base model.
- Implements the GRPO algorithm and adopts rule-based rewards, making multimodal large model training stable.
- Trained on a dataset of 50,000 mathematical geometry reasoning examples.
- Evaluated on three commonly used mathematical benchmarks: MathVista (Testmini), MathVerse (Testmini & Vision_Only), and WeMath. Scores for other models come from OpenCompass.
How to Use The repository page includes model weight downloads, inference code examples, and training process explanations. The inference code section provides code snippets on how to use the chat model, but the materials do not give the specific code content. Installation commands, dependencies, and running steps are not yet provided.
Applicable Scenarios Suitable for research and applications in multimodal tasks related to mathematical geometry reasoning.
Limitations The materials do not specify the license, details beyond the model parameter scale, hardware requirements, inference latency, fees, or pricing. Some content (such as specific evaluation scores and complete code) is not yet provided.
