What is 360gpt2-o1
360gpt2-o1 is an AI large model independently developed by 360, with significant improvements in reasoning ability, especially outstanding performance in mathematical and logical reasoning tasks. The model achieves technological breakthroughs through synthetic data optimization, model post-training, and the "slow thinking" paradigm, and has achieved excellent results in multiple authoritative evaluations. In basic mathematics evaluations (such as MATH, Gaokao Mathematics) and authoritative mathematics competitions (including AIME24, AMC23), 360gpt2-o1 surpasses the previous-generation model 360gpt2-pro and outperforms GPT-4o model. In mathematics competition evaluations, 360gpt2-o1 surpasses Alibaba's latest open-source o1 series model QWQ-32B-preview。
Main features of 360gpt2-o1
- Improved reasoning ability: 360gpt2-o1 performs excellently on mathematical and logical reasoning tasks, especially with significant improvements in reasoning ability.
- Synthetic data optimization: Through methods such as instruction synthesis and quality/diversity filtering, it solves the problem of scarcity of high-quality mathematical and logical reasoning data and effectively expands the training dataset.
- Model post-training: It adopts a two-stage training strategy, first using a small model to generate diverse reasoning paths, then using a large model for RFT training and reinforcement learning training, improving the model's reasoning ability and reflective error-correction ability.
- "Slow thinking" paradigm: Based on Monte Carlo tree search to explore diverse solutions, it introduces an LLM for error verification and correction, simulating the process of human step-by-step reasoning and reflection, ultimately forming a long chain of thought that includes reflection, verification, error correction, and backtracking.
Technical principles of 360gpt2-o1
- Data synthesis and filtering: Through synthetic data optimization, 360gpt2-o1 can generate and filter high-quality training data, which is crucial for model training.
- Two-stage training strategy: The first stage uses a small model to generate reasoning paths, and the second stage uses a large model for training, enabling the model to improve the accuracy and depth of reasoning while maintaining reasoning diversity.
- Combination of Monte Carlo tree search and LLM: Through Monte Carlo tree search, the model can explore multiple possible solutions, and the introduction of the LLM provides the model with error verification and correction capabilities, enhancing the model's robustness.
How to use 360gpt2-o1
- Visit 360 Brain: Currently 360gpt2-o1 has been launched on360 BrainAPI open platform.
- Experience address：https://ai.360.com/playground/?model=360gpt2-o1?src=weixinmp
Application scenarios of 360gpt2-o1
- Mathematical problem solving: 360gpt2-o1 has achieved remarkable results in basic mathematics evaluations (such as MATH, Gaokao Mathematics) and authoritative mathematics competitions (including AIME24, AMC23), demonstrating strong capabilities in mathematical problem solving.
- Logical reasoning: Through "slow thinking" technology, the model simulates the process of human step-by-step reasoning and reflection, and has the ability to solve complex logical problems.
- Programming problems: In performance in fields such as mathematics and programming, it approaches or even surpasses o1, and 360gpt2-o1 provides support for programming problem solving.
- Complex problem solving: 360gpt2-o1 can handle complex problems that require deep-level logical reasoning ability, including the ability for self-reflection and error correction.
- Education and academia: The application of models to mathematical and logical problems in the field of education can assist teaching and academic research.
- Enterprise decision support: Through logical reasoning and data analysis, 360gpt2-o1 can assist enterprises in providing logical support during complex decision-making processes.