Alibaba International Digital Commerce, the web e-commerce arm of Alibaba, has released the artificial intelligence model LLM under the name Marco-o1. This model has shown a significant increase in math problem-solving accuracy with an impressive 90% success rate, despite its compact size and incremental thinking approach.
Marco-o1 is constructed from Qwen2-7B but incorporates 4 techniques to enhance its performance. These include training with the chain-of-thought dataset, where a synthetic dataset helps analyze problems before answering, and utilizing Monte Carlo Tree Search (MCTS) to generate multiple possible solutions and select the best one for each step of the process.
The results of these enhancements have led to a substantial improvement in MGSM test outcomes, particularly noticeable when translating sentences from Chinese to English. The team behind Marco-o1 acknowledges its development following the footsteps of OpenAI o1 but in a much smaller scale. This report focuses solely on the improved performance of MGSM tests without other evaluation metrics. It is intriguing to note the proximity of this model’s release to the release of the QwQ model by the Qwen team.
Source: HuggingFace: AIDC-AI
TLDR: Alibaba’s Marco-o1 AI model demonstrates impressive math problem-solving capabilities, boasting a 90% success rate. Using innovative techniques and datasets, the model outperforms previous iterations and shows promise for future advancements in AI technology.
Leave a Comment