|
--- |
|
license: llama2 |
|
datasets: |
|
- gair-prox/open-web-math-pro |
|
language: |
|
- en |
|
pipeline_tag: text-generation |
|
tags: |
|
- llama2 |
|
- math |
|
- reasoning |
|
base_model: |
|
- meta-llama/Llama-2-7b-hf |
|
--- |
|
|
|
# Llama-2-7B-ProXMath |
|
|
|
<p align="center"> |
|
<img src="prox-teaser.png"> |
|
</p> |
|
|
|
[ArXiv](https://arxiv.org/abs/2409.17115) | [Data: OpenWebMath-Pro](https://huggingface.co/datasets/gair-prox/open-web-math-pro) | [Code](https://github.com/GAIR-NLP/program-every-example) |
|
|
|
**Llama-2-7B-ProXMath** is a math-adapted Llama-2-7B model that is continually pre-trained on [OpenWebMath-Pro](https://huggingface.co/datasets/gair-prox/open-web-math-pro) (a refined version by ProX) for **10**B tokens. |
|
|
|
## Evaluations |
|
|
|
ProX models are evaluated on 9 common math reasoning benchmarks. |
|
|
|
| Model | asdiv | gsm8k | mathqa | mawps | minerva_math | mmlu_stem | sat_math | svamp | tabmwp | average | |
|
|---------------------|:-----:|:-----:|:------:|:-----:|:------------:|:---------:|:--------:|:-----:|:------:|:-------:| |
|
| Llama-2-7B | 51.6 | 14.1 | 12.5 | 63.6 | 3.8 | 32.9 | 34.4 | 39.5 | 30.9 | 31.48 | |
|
| Llama-2-7B-ProXMath | 63.7 | 30.6 | 40.1 | 79.3 | 16.8 | 43.8 | 53.1 | 50.2 | 37.3 | 46.1 | |
|
|
|
### Citation |
|
``` |
|
@article{zhou2024programming, |
|
title={Programming Every Example: Lifting Pre-training Data Quality like Experts at Scale}, |
|
author={Zhou, Fan and Wang, Zengzhi and Liu, Qian and Li, Junlong and Liu, Pengfei}, |
|
journal={arXiv preprint arXiv:2409.17115}, |
|
year={2024} |
|
} |
|
``` |
|
|