---
library_name: transformers
license: apache-2.0
datasets:
- monology/pile-uncopyrighted
- MiniLLM/pile-diff_samp-qwen_1.8B-qwen_104M-r0.5
language:
- en
metrics:
- accuracy
pipeline_tag: text-generation
---

# MiniPLM-llama3.1-212M

[paper](https://arxiv.org/abs/2410.17215) | [code](https://github.com/thu-coai/MiniPLM)

**MiniPLM-llama3.1-212M** is a 212M model with the [LLaMA3.1 achitecture](https://arxiv.org/abs/2407.21783) pre-trained from scratch on [the Pile](https://huggingface.co./datasets/monology/pile-uncopyrighted) using the MiniPLM knowledge distillation framework with the [offcial Qwen1.5-1.8B](https://huggingface.co./Qwen/Qwen1.5-1.8B) as the teacher model.
This model shows the flexibility of the MiniPLM framework in conducting knowledge distillation across model families.

We also open-source the [pre-training corpus](https://huggingface.co./datasets/MiniLLM/pile-diff_samp-qwen_1.8B-qwen_104M-r0.5) refined by Difference Sampling in MiniPLM for reproducibility.

<p align='left'>
    <img src="https://cdn-uploads.huggingface.co/production/uploads/624ac662102fcdff87be51b9/2BqT0NgkmIXYlktovw9kG.png" width="1000">
</p>

## Evaluation

MiniPLM models achieves better performance given the same computation and scales well across model sizes:

<p align='left'>
    <img src="https://cdn-uploads.huggingface.co/production/uploads/624ac662102fcdff87be51b9/EOYzajQcwQFT5PobqL3j0.png" width="1000">
</p>

## Baseline Models
+ [Conventional Pre-Training](https://huggingface.co./MiniLLM/Pretrain-LLama3.1-130M)

## Citation

```bibtex
@article{miniplm,
    title={MiniPLM: Knowledge Distillation for Pre-Training Language Models}, 
    author={Yuxian Gu and Hao Zhou and Fandong Meng and Jie Zhou and Minlie Huang},
    journal={arXiv preprint arXiv:2410.17215},
    year={2024}
}
```