| ArXiv | Models | Data | Code |
Mix of Minimal Optimal Sets (MMOS) of dataset has two advantages for two aspects, higher performance and lower construction costs on math reasoning.
Model | Size | GSM8K | SVAMP | ASDiv | MATH | Size | GSM8K | SVAMP | ASDiv | MATH | Size | GSM8K | SVAMP | ASDiv | MATH |
---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
WizardMath | 7B | 54.9 | 57.3 | 59.1 | 10.7 | 13B | 63.9 | 64.3 | 65.8 | 14.0 | 34B | - | - | - | - |
MAMMOTH | 7B | 53.6 | 67.7 | 31.5 | - | 13B | 62.0 | 72.4 | - | 34.2 | 34B | - | - | - | - |
MetaMath | 7B | 66.5 | - | - | 19.8 | 13B | 72.3 | - | - | 22.4 | 34B | - | - | - | - |
MathCoder-L | 7B | 64.2 | 71.5 | - | 23.3 | 13B | 72.6 | 76.9 | - | 29.9 | 34B | - | - | - | - |
MathCoder-CL | 7B | 67.8 | 70.7 | - | 30.2 | 13B | 74.1 | 78.0 | - | 35.9 | 34B | - | - | - | - |
TORA | 7B | 68.8 | 68.2 | 73.9 | 40.1 | 13B | 72.7 | 72.9 | 77.2 | 43.0 | 34B | - | - | - | - |
TORA-CODE | 7B | 72.6 | 70.4 | 78.7 | 44.6 | 13B | 75.8 | 75.7 | 81.4 | 48.1 | 34B | 80.7 | 80.5 | 84.2 | 50.8 |
MMOS | 7B | 69.9 | 73.4 | 76.8 | 40.2 | 13B | 74.8 | 77.0 | 80.0 | 43.2 | 34B | - | - | - | - |
MMOS-CODE | 7B | 73.9 | 76.4 | 78.6 | 44.3 | 13B | 77.1 | 77.5 | 81.9 | 48.1 | 34B | 81.7 | 81.9 | 82.8 | 48.8 |
MMOS-MinCODE | 7B | 70.3 | 72.5 | 76.7 | 44.6 | 13B | - | - | - | - | 34B | - | - | - | - |
MMOS-LLEMMA | 7B | 76.5 | 77.7 | 81.4 | 48.8 | 13B | - | - | - | - | 34B | 82.8 | 81.8 | 84.8 | 51.3 |
MMOS-DeepSeekMath | 7B | 80.5 | 79.3 | 87.6 | 55.0 | 13B | - | - | - | - | 34B | - | - | - | - |
MMOS-DeepSeekMath(SC,k=50) | 7B | 87.2 | - | - | 63.7 | 13B | - | - | - | - | 34B | - | - | - | - |
git clone https://github.com/cyzhh/MMOS.git
cd MMOS
conda create -n MMOS python=3.10
conda activate MMOS
pip install -r requirements.txt
To identify the minimal optimal set, we follow these steps: 1) Sample a sufficient number of correct reasoning paths to form initial set. 2) Implement a deduplication algorithm to obtain its deduplicated subset. 3) Conduct a statistical analysis on the upper limit of reasoning paths per question k with the subset data amount N. 4) Perform SFT on several subsets to analyze the impact of removing duplicates and keeping varied reasoning paths.
We use ToRA series to generate QA-pairs from open source dataset GSM8K, MATH, TAL-SCQ. The QA-pairs are processed by our deduplication algorithm, resulting in the dataset MMOS
. The total number of QA-pairs is 135K.
The DATA, which we publish at π HuggingFace, need to be placed under the relative path, ./train_data/MMOS/
.
If you are interested in our work, we will publish details about the data processing aspects after the paper is published.
Following scripts/generate.sh
:
You can generate a data set for testing the numerical robustness of model performance by executing the following script commandοΌ
bash scripts/generate.sh
bash scripts/attack.sh
bash scripts/rerank.sh
Due to resource constraints, we performed supervised fine-tuning on CodeLLaMA 7B, CodeLLaMA 13B and CodeLLaMA 34B using our dataset on A100 40G GPUs. To reproduce our work from CodeLLaMA 7B/13B, you can train according to the following instruction. You can also train the 34B model through DDP script instructions.
bash scripts/train_single.sh codellama 7b
bash scripts/train.sh codellama 34b
bash scripts/infer.sh
If you find this repository helpful, please consider citing our paper:
@misc{chen2024empirical,
title={An Empirical Study of Data Ability Boundary in LLMs' Math Reasoning},
author={Zui Chen and Yezeng Chen and Jiaqi Han and Zhijie Huang and Ji Qi and Yi Zhou},
year={2024},
eprint={2403.00799},
archivePrefix={arXiv},
primaryClass={cs.CL}
}