Summary of Common 7b Language Models Already Possess Strong Math Capabilities, by Chen Li et al.
Common 7B Language Models Already Possess Strong Math Capabilities
by Chen Li, Weiqi Wang, Jingcheng Hu, Yixuan Wei, Nanning Zheng, Han Hu, Zheng Zhang, Houwen Peng
First submitted to arxiv on: 7 Mar 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 paper explores the mathematical capabilities of language models, specifically the LLaMA-2 7B model with common pre-training. Contrary to previous assumptions, this model demonstrates strong math abilities, achieving impressive accuracy on GSM8K (97.7%) and MATH (72.0%) benchmarks when generating multiple responses. However, the model’s primary limitation is its inconsistency in eliciting mathematical capabilities. Scaling up SFT data enhances reliability, but is constrained by the scarcity of public math questions. To overcome this, the authors employ synthetic data, which proves nearly as effective as real data and shows no saturation when scaled up to one million samples. This approach achieves higher accuracy on GSM8K (82.6%) and MATH (40.6%) benchmarks compared to previous models. |
| Low | GrooveSquid.com (original content) | Low Difficulty Summary This paper is about a new type of computer model that can do math problems really well. The model, called LLaMA-2 7B, was trained using common language, not just math formulas. It’s able to solve complex math problems with high accuracy, but it has trouble consistently doing this. To fix this, the researchers tried increasing the amount of training data and found that this helped a lot. They also created fake math questions, which allowed them to train the model even further. This approach led to even better results than previous models. |
Keywords
» Artificial intelligence » Llama » Synthetic data
