README.md

---
license: apache-2.0
datasets:
- togethercomputer/RedPajama-Data-1T
language:
- en
pipeline_tag: text-generation
library_name: transformers
---

## PDS-160M

[paper](https://arxiv.org/abs/2410.07064) | [code](https://github.com/microsoft/LMOps/tree/main/data_selection)

**PDS-160M** is a 160M model with [Mistral](https://arxiv.org/abs/2310.06825) achitecture pre-trained from scratch on the data selected from the CC split of [Redpajama](https://github.com/togethercomputer/RedPajama-Data), using the PDS framework.

The PDS framework is based on the [Pontryagin's maximum principle](https://en.wikipedia.org/wiki/Pontryagin%27s_maximum_principle#:~:text=Pontryagin's%20maximum%20principle%20is%20used,the%20state%20or%20input%20controls.) for optimal pre-training data selection, which not only enjoy strong theoretical support but is also scalable for training large language models. 

Please refer to our [paper](https://arxiv.org/abs/2410.07064) for more details.

### Overview of the theory:

<p align='left'>
    <img src="https://cdn-uploads.huggingface.co/production/uploads/624ac662102fcdff87be51b9/Hdw83Vsb305GRlsqB7c34.png" width="700">
</p>

### Overview of the PDS framework:

<p align='left'>
    <img src="https://cdn-uploads.huggingface.co/production/uploads/624ac662102fcdff87be51b9/YPwluLyZGK7DACH1WqDUN.png" width="700">
</p>

### Evaluation

PDS-selected data improves the performance of language models pre-trained from scratch and saves pre-training comptation. The improvement scales up to large model sizes.

<p align='left'>
    <img src="https://cdn-uploads.huggingface.co/production/uploads/624ac662102fcdff87be51b9/6undIr37d10qD73TDiPDK.png" width="600">
</p>

### Baseline

[Conventional Pre-training](https://huggingface.co/Data-Selection/BSL-160M)

### Citation

```bibtex
@article{gu2024data,
  title={Data Selection via Optimal Control for Language Models},
  author={Gu, Yuxian and Dong, Li and Wang, Hongning and Hao, Yaru and Dong, Qingxiu and Wei, Furu and Huang, Minlie},
  journal={arXiv preprint arXiv:2410.07064},
  year={2024}
}
```