|
|
--- |
|
|
datasets: |
|
|
- monology/pile-uncopyrighted |
|
|
- MiniLLM/pile-tokenized |
|
|
language: |
|
|
- en |
|
|
library_name: transformers |
|
|
license: apache-2.0 |
|
|
metrics: |
|
|
- accuracy |
|
|
pipeline_tag: text-generation |
|
|
--- |
|
|
|
|
|
# Ref-Pretrain-Qwen-104M |
|
|
|
|
|
[paper](https://arxiv.org/abs/2410.17215) | [code](https://github.com/thu-coai/MiniPLM) |
|
|
|
|
|
**Ref-Pretrain-Qwen-104M** is a 104M model with Qwen achitecture conventionally pre-trained from scratch on [the Pile](https://huggingface.co/datasets/monology/pile-uncopyrighted) for 5B tokens. |
|
|
|
|
|
We also open-source the tokenized [pre-training corpus](https://huggingface.co/datasets/MiniLLM/pile-tokenized) for reproducibility. |
|
|
|
|
|
**It is used as the reference model in the MiniPLM knwoledge distillation framework to construct the [refined pre-training corpus](https://huggingface.co/datasets/MiniLLM/pile-diff_samp-qwen_1.8B-qwen_104M-r0.5).** |
|
|
**The data is then used to train [MiniPLM models](https://huggingface.co/collections/MiniLLM/miniplm-6712c0fdf09ef7e8da7d39bd).** |
|
|
|
|
|
## 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> |
|
|
|
|
|
## Citation |
|
|
|
|
|
```bibtext |
|
|
@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} |
|
|
} |
|
|
``` |
