一、 个人在openwebtext数据集上训练得到的electra-small模型
二、 复现结果(dev dataset)
| Model | CoLA | SST | MRPC | STS | QQP | MNLI | QNLI | RTE | Avg. |
|---|---|---|---|---|---|---|---|---|---|
| Metrics | MCC | Acc | Acc | Spearman | Acc | Acc | Acc | Acc | |
| 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) | 55.82 | 89.67 | 87.0 | 86.96 | 89.28 | 80.08 | 87.50 | 66.07 | 80.30 |
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 62.5W
- GPU RTX3090
- 训练时间总共耗费2.5天
四、 使用
import torch
from transformers.models.electra import ElectraModel, ElectraTokenizer
tokenizer = ElectraTokenizer.from_pretrained("junnyu/electra_small_discriminator")
model = ElectraModel.from_pretrained("junnyu/electra_small_discriminator")
inputs = tokenizer("Beijing is the capital of China.", return_tensors="pt")
with torch.no_grad():
outputs = model(**inputs)
print(outputs[0].shape)
