Summary of Jiuzhang3.0: Efficiently Improving Mathematical Reasoning by Training Small Data Synthesis Models, By Kun Zhou et al.
JiuZhang3.0: Efficiently Improving Mathematical Reasoning by Training Small Data Synthesis Models
by Kun Zhou, Beichen Zhang, Jiapeng Wang, Zhipeng Chen, Wayne Xin Zhao, Jing Sha, Zhichao Sheng, Shijin Wang, Ji-Rong Wen
First submitted to arxiv on: 23 May 2024
Categories
- Main: Computation and Language (cs.CL)
- Secondary: Artificial Intelligence (cs.AI)
GrooveSquid.com Paper Summaries
GrooveSquid.com’s goal is to make artificial intelligence research accessible by summarizing AI papers in simpler terms. Each summary below covers the same AI paper, written at different levels of difficulty. The medium difficulty and low difficulty versions are original summaries written by GrooveSquid.com, while the high difficulty version is the paper’s original abstract. Feel free to learn from the version that suits you best!
| Summary difficulty | Written by | Summary |
|---|---|---|
| High | Paper authors | High Difficulty Summary Read the original abstract here |
| Medium | GrooveSquid.com (original content) | Medium Difficulty Summary The proposed method trains a small large language model (LLM) to efficiently generate high-quality pre-training data for mathematical reasoning tasks. This is achieved by leveraging GPT-4 to synthesize math problems, then distilling the synthesis capability into the small LLM. The approach uses prompts based on human education stages and gradient-based influence estimation to select valuable math-related texts. The resulting dataset is used to train the JiuZhang3.0 model, which achieves state-of-the-art performance on several mathematical reasoning datasets. |
| Low | GrooveSquid.com (original content) | Low Difficulty Summary The researchers developed a way to make large language models better at doing math problems. They did this by using an even bigger language model (GPT-4) to create lots of math problems, then teaching a smaller model how to do those problems too. The small model only needs to use the GPT-4’s ideas 9,300 times and learn from just 4.6 billion pieces of information. This makes it much faster and cheaper than training the big model itself. The small model is really good at doing math problems now, even better than some other models that were trained in a different way. |
Keywords
» Artificial intelligence » Gpt » Language model » Large language model
