|
--- |
|
license: apache-2.0 |
|
datasets: |
|
- gair-prox/RedPajama-pro |
|
language: |
|
- en |
|
base_model: |
|
- gair-prox/RedPJ-ProX-0.3B |
|
pipeline_tag: text-generation |
|
library_name: transformers |
|
tags: |
|
- llama |
|
- code |
|
--- |
|
|
|
# Math-chunk-refining-lm |
|
|
|
<p align="center"> |
|
<img src="prox-teaser.png"> |
|
</p> |
|
|
|
[ArXiv](http://arxiv.org/abs/2409.17115) | [Code](https://github.com/GAIR-NLP/program-every-example) |
|
|
|
**Math-chunk-refining-lm** is an adapted [0.3B-ProX](https://huggingface.co./gair-prox/RedPJ-ProX-0.3B) model, fine-tuned for doc level refining via program generation, and can be applied over math pre-training corpus such as open-web-math. |
|
|
|
<p align="center"> |
|
<img src="func_design.png"> |
|
</p> |
|
|
|
### 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} |
|
} |
|
``` |