---
license: apache-2.0
datasets:
- gair-prox/FineWeb-pro
language:
- en
tags:
- llama
pipeline_tag: text-generation
library_name: transformers
---

# FW-ProX-1.7B

<p align="center">
  <img src="prox-teaser.png">
</p>

[ArXiv](https://arxiv.org/abs/2409.17115) | [Models](https://huggingface.co/gair-prox/FW-ProX-1.7B) | [Data](https://huggingface.co/datasets/gair-prox/FineWeb-pro) | [Code](https://github.com/GAIR-NLP/program-every-example)

**FW-ProX-1.7B** is a small language model. It was and trained on the [FineWeb-pro](https://huggingface.co/datasets/gair-prox/FineWeb-pro) for 50B tokens. 

## Evaluations

ProX models are evaluated over 10 language model benchmarks in zero-shot setting.

|                       | ArC-c | ARC-e | CSQA  | HellaS | MMLU  | OBQA  | PiQA  | SIQA  | WinoG | SciQ  | AVG  |
|-----------------------|-------|-------|-------|-----------|-------|-------|-------|-------|-------|-------|------|
| raw | 28.5 | 52.6 | 33.9 | 53.2 | 29.8 | 32.6 | 72.9 | 40.2 | 53.0 | 77.1 | 47.4 |
| ours | 34.4 | 63.9 | 32.6 | 53.0 | 33.1 | 34.4 | 73.1 | 39.3 | 52.7 | 81.5 | 49.8 |

### 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}
}
```