MetaMathQA

View the project page:
https://meta-math.github.io/
see our paper at https://arxiv.org/abs/2309.12284

## Note

All MetaMathQA data are augmented from the training sets of GSM8K and MATH.
<span style="color:red"><b>None of the augmented data is from the testing set.</b></span>

You can check the `original_question` in `meta-math/MetaMathQA`, each item is from the GSM8K or MATH train set.

## Model Details

MetaMath-Mistral-7B is fully fine-tuned on the MetaMathQA datasets and based on the powerful Mistral-7B model. It is glad to see using MetaMathQA datasets and changing the base model from llama-2-7B to Mistral-7b can boost the GSM8K performance from 66.5 to **77.7**.

To fine-tune Mistral-7B, I would suggest using a smaller learning rate (usually 1/5 to 1/10 of the lr for LlaMa-2-7B) and staying other training args unchanged.
More training details and scripts can be seen at [https://github.com/meta-math/MetaMath](https://github.com/meta-math/MetaMath).

## Installation

```
pip install transformers==4.35.0
pip install torch==2.0.1
pip install sentencepiece==0.1.99
pip install tokenizers==0.13.3
pip install accelerate==0.21.0
pip install bitsandbytes==0.40.0
pip install vllm
pip install fraction
pip install protobuf
```

## Model Usage

prompting template:

'''

"Below is an instruction that describes a task. "
"Write a response that appropriately completes the request.\n\n"
"### Instruction:\n{instruction}\n\n### Response: Let's think step by step."

'''

where you need to use your query question to replace the {instruction}

There is another interesting repo about Arithmo-Mistral-7B at [https://huggingface.co/akjindal53244/Arithmo-Mistral-7B](https://huggingface.co/akjindal53244/Arithmo-Mistral-7B), where they combine our MetaMathQA dataset and MathInstruct datasets to train a powerful model. Thanks agian for their contributions.
We would also try to train the combination of **MetaMathQA** and **MathInstruct** datasets, and also open all the results and training details.

## Experiments

| Model | GSM8k Pass@1 | MATH Pass@1 |
|---------------------|--------------|-------------|
| MPT-7B | 6.8 | 3.0 |
| Falcon-7B | 6.8 | 2.3 |
| LLaMA-1-7B | 11.0 | 2.9 |
| LLaMA-2-7B | 14.6 | 2.5 |
| MPT-30B | 15.2 | 3.1 |
| LLaMA-1-13B | 17.8 | 3.9 |
| GPT-Neo-2.7B | 19.5 | -- |
| Falcon-40B | 19.6 | 2.5 |
| Baichuan-chat-13B | 23.9 | -- |
| Vicuna-v1.3-13B | 27.6 | -- |
| LLaMA-2-13B | 28.7 | 3.9 |
| InternLM-7B | 31.2 | -- |
| ChatGLM-2-6B | 32.4 | -- |
| GPT-J-6B | 34.9 | -- |
| LLaMA-1-33B | 35.6 | 3.9 |
| LLaMA-2-34B | 42.2 | 6.24 |
| RFT-7B | 50.3 | -- |
| LLaMA-1-65B | 50.9 | 10.6 |
| Qwen-7B | 51.6 | -- |
| WizardMath-7B | 54.9 | 10.7 |
| LLaMA-2-70B | 56.8 | 13.5 |
| WizardMath-13B | 63.9 | 14.0 |
| MAmmoTH-7B (COT) | 50.5 | 10.4 |
| MAmmoTH-7B (POT+COT)| 53.6 | 31.5 |
| Arithmo-Mistral-7B | 74.7 | 25.3 |
| MetaMath-7B | 66.5 | 19.8 |
| MetaMath-13B | 72.3 | 22.4 |
| 🔥 **MetaMath-Mistral-7B** | **77.7** | **28.2** |

We encourage anyone to use our MetaMathQA datasets. We are very happy to see the following models trained by MetaMathQA achieve a very promising performance!

OpenChat-3.5 (https://huggingface.co/openchat/openchat_3.5)

CausalLM (https://huggingface.co/CausalLM/14B)

zephyr (https://huggingface.co/qblocks/zephyr-7b-alpha_metamathqa)

Ziya2 (https://huggingface.co/IDEA-CCNL/Ziya2-13B-Base)

# Citation

```bibtex
@article{yu2023metamath,
title={MetaMath: Bootstrap Your Own Mathematical Questions for Large Language Models},
author={Yu, Longhui and Jiang, Weisen and Shi, Han and Yu, Jincheng and Liu, Zhengying and Zhang, Yu and Kwok, James T and Li, Zhenguo and Weller, Adrian and Liu, Weiyang},
journal={arXiv preprint arXiv:2309.12284},
year={2023}
}
```

访问项目主页：https://meta-math.github.io/
查阅我们的论文：https://arxiv.org/abs/2309.12284

## 备注

所有MetaMathQA数据集均源自GSM8K与MATH的训练集，经数据增强生成。
<span style="color:red"><b>所有增强数据均未使用测试集样本。</b></span>

您可在`meta-math/MetaMathQA`中查看`original_question`字段，每条原始问题均来自GSM8K或MATH的训练集。

## 模型详情

MetaMath-Mistral-7B基于高性能Mistral-7B模型，在MetaMathQA数据集上完成全量微调。我们观察到，将基座模型从LLaMA-2-7B替换为Mistral-7B并使用MetaMathQA数据集进行微调后，模型在GSM8K数据集上的Pass@1准确率从66.5提升至**77.7**。

针对Mistral-7B的微调，建议使用更小的学习率（通常为LLaMA-2-7B微调学习率的1/5至1/10），其余训练参数保持不变即可。
更多训练细节与脚本可参考：[https://github.com/meta-math/MetaMath](https://github.com/meta-math/MetaMath)。

## 环境安装

pip install transformers==4.35.0
pip install torch==2.0.1
pip install sentencepiece==0.1.99
pip install tokenizers==0.13.3
pip install accelerate==0.21.0
pip install bitsandbytes==0.40.0
pip install vllm
pip install fraction
pip install protobuf

## 模型使用

### 提示模板

"以下是一项描述具体任务的指令。"
"请编写一段恰当的回复以完成该请求。

"
"### 指令：
{instruction}

### 回复：我们一步步来思考。"

其中需将您的查询问题替换`{instruction}`占位符。

另有一项关于Arithmo-Mistral-7B的开源项目：[https://huggingface.co/akjindal53244/Arithmo-Mistral-7B](https://huggingface.co/akjindal53244/Arithmo-Mistral-7B)，该项目将我们的MetaMathQA数据集与MathInstruct数据集结合，训练得到了高性能模型。在此感谢其团队的贡献。

我们也计划结合**MetaMathQA**与**MathInstruct**数据集开展训练，并公开所有实验结果与训练细节。

## 实验结果

| 模型 | GSM8K Pass@1 准确率 | MATH Pass@1 准确率 |
|---------------------|--------------|-------------|
| MPT-7B | 6.8 | 3.0 |
| Falcon-7B | 6.8 | 2.3 |
| LLaMA-1-7B | 11.0 | 2.9 |
| LLaMA-2-7B | 14.6 | 2.5 |
| MPT-30B | 15.2 | 3.1 |
| LLaMA-1-13B | 17.8 | 3.9 |
| GPT-Neo-2.7B | 19.5 | -- |
| Falcon-40B | 19.6 | 2.5 |
| Baichuan-chat-13B | 23.9 | -- |
| Vicuna-v1.3-13B | 27.6 | -- |
| LLaMA-2-13B | 28.7 | 3.9 |
| InternLM-7B | 31.2 | -- |
| ChatGLM-2-6B | 32.4 | -- |
| GPT-J-6B | 34.9 | -- |
| LLaMA-1-33B | 35.6 | 3.9 |
| LLaMA-2-34B | 42.2 | 6.24 |
| RFT-7B | 50.3 | -- |
| LLaMA-1-65B | 50.9 | 10.6 |
| Qwen-7B | 51.6 | -- |
| WizardMath-7B | 54.9 | 10.7 |
| LLaMA-2-70B | 56.8 | 13.5 |
| WizardMath-13B | 63.9 | 14.0 |
| MAmmoTH-7B (COT) | 50.5 | 10.4 |
| MAmmoTH-7B (POT+COT)| 53.6 | 31.5 |
| Arithmo-Mistral-7B | 74.7 | 25.3 |
| MetaMath-7B | 66.5 | 19.8 |
| MetaMath-13B | 72.3 | 22.4 |
| 🔥 **MetaMath-Mistral-7B** | **77.7** | **28.2** |

我们鼓励所有研究者使用MetaMathQA数据集，同时欣喜地看到以下基于MetaMathQA数据集训练的模型取得了优异的性能表现：
- OpenChat-3.5（https://huggingface.co/openchat/openchat_3.5）
- CausalLM（https://huggingface.co/CausalLM/14B）
- zephyr（https://huggingface.co/qblocks/zephyr-7b-alpha_metamathqa）
- Ziya2（https://huggingface.co/IDEA-CCNL/Ziya2-13B-Base）

## 引用

bibtex
@article{yu2023metamath,
title={MetaMath: Bootstrap Your Own Mathematical Questions for Large Language Models},
author={Yu, Longhui and Jiang, Weisen and Shi, Han and Yu, Jincheng and Liu, Zhengying and Zhang, Yu and Kwok, James T and Li, Zhenguo and Weller, Adrian and Liu, Weiyang},
journal={arXiv preprint arXiv:2309.12284},
year={2023}
}

论文标题可译为：《MetaMath：为大语言模型（Large Language Model）自主生成数学问题》。