README.md

---
language: "en"
thumbnail: "https://github.com/junnyu"
tags:
- pytorch
- electra
- roformer
- rotary position embedding
license: "MIT"
datasets:
- openwebtext

---
# 一、 个人在openwebtext数据集上添加rotary-position-embedding，训练得到的electra-small模型

# 二、 复现结果(dev dataset)
|Model|CoLA|SST|MRPC|STS|QQP|MNLI|QNLI|RTE|Avg.|
|---|---|---|---|---|---|---|---|---|---|
|ELECTRA-Small-OWT(original)|56.8|88.3|87.4|86.8|88.3|78.9|87.9|68.5|80.36|
|**ELECTRA-Small-OWT (this)**| xx|xx|xx|xx|xx|xx|xx|xx|xx|

# 三、 训练细节
- 数据集 openwebtext
- 训练batch_size 256
- 学习率lr  5e-4
- 最大句子长度max_seqlen  128
- 训练total step  50W
- GPU RTX3090
- 训练时间总共耗费55h

### wandb日志 
https://wandb.ai/junyu/electra_rotary_small_pretrain?workspace=user-junyu

# 四、安装
```bash
pip install roformer
或
pip install git+https://github.com/JunnYu/RoFormer_pytorch.git
```

# 五、 使用
```python
import torch
from roformer import RoFormerModel
from transformers import ElectraTokenizer
tokenizer = ElectraTokenizer.from_pretrained("junnyu/roformer_small_discriminator")
model = RoFormerModel.from_pretrained("junnyu/roformer_small_discriminator")
inputs = tokenizer("Beijing is the capital of China.", return_tensors="pt")
with torch.no_grad():
    outputs = model(**inputs)
    print(outputs[0].shape)
```