Mathoctopus/Parallel_xRFT_13B: Multilingual Math Reasoning
Mathoctopus/Parallel_xRFT_13B is a 13 billion parameter model from the MathOctopus series, built upon the LLaMA 2 architecture. Developed by Mathoctopus, this model is specifically engineered to excel in multilingual mathematical problem-solving. It leverages a unique parallel-training strategy enhanced with multilingual rejection sampling (xRFT).
Key Capabilities
- Multilingual Math Problem Solving: Trained on the extensive MGSM8KInstruct Dataset, which includes 73.6K problems across ten languages (English, Swahili, Chinese, Bengali, German, Spanish, French, Japanese, Russian, Thai).
- Enhanced Performance: Demonstrates superior performance in few-shot scenarios for multilingual math reasoning compared to conventional open-source LLMs and ChatGPT.
- Rejection Sampling: Incorporates multilingual rejection sampling (xRFT) during training to further refine its mathematical reasoning abilities.
Good For
- Educational Software: Ideal for applications requiring accurate solutions to math problems in various languages.
- Tutoring Systems: Can be integrated into AI-powered tutoring platforms to assist users with multilingual mathematical queries.
- Research in Multilingual LLMs: Useful for researchers exploring the frontiers of language models in complex reasoning tasks across diverse linguistic contexts. The model achieved an overall score of 47.6% on MGSM8K and 44.5% on MSVAMP for the 13B xRFT-Parallel variant.