---
language: 
- zh
tags:
- chatglm
- pytorch
- zh
- Text2Text-Generation
license: "apache-2.0"
widget:
- text: "对下面中文拼写纠错：\n少先队员因该为老人让坐。\n答："

---

# Chinese Spelling Correction LoRA Model
ChatGLM3-6B中文纠错LoRA模型

`shibing624/chatglm3-6b-csc-chinese-lora` evaluate test data：

The overall performance of shibing624/chatglm3-6b-csc-chinese-lora on CSC **test**:

|prefix|input_text|target_text|pred|
|:-- |:--- |:--- |:-- |
|对下面文本纠错：|少先队员因该为老人让坐。|少先队员应该为老人让座。|少先队员应该为老人让座。|

在CSC测试集上生成结果纠错准确率高，由于是基于ChatGLM3-6B模型，结果常常能带给人惊喜，不仅能纠错，还带有句子润色和改写功能。


## Usage

本项目开源在 pycorrector 项目：[textgen](https://github.com/shibing624/pycorrector)，可支持ChatGLM原生模型和LoRA微调后的模型，通过如下命令调用：

Install package:
```shell
pip install -U pycorrector
```

```python
from pycorrector.gpt.gpt_model import GptModel
model = GptModel("chatglm", "THUDM/chatglm3-6b", peft_name="shibing624/chatglm3-6b-csc-chinese-lora")
r = model.predict(["对下面文本纠错：\n少先队员因该为老人让坐。"])
print(r) # ['少先队员应该为老人让座。']
```

## Usage (HuggingFace Transformers)
Without [pycorrector](https://github.com/shibing624/pycorrector), you can use the model like this: 

First, you pass your input through the transformer model, then you get the generated sentence.

Install package:
```
pip install transformers 
```

```python
import sys
from peft import PeftModel
from transformers import AutoModel, AutoTokenizer

sys.path.append('..')

model = AutoModel.from_pretrained("THUDM/chatglm3-6b", trust_remote_code=True, device_map='auto')
model = PeftModel.from_pretrained(model, "shibing624/chatglm3-6b-csc-chinese-lora")
model = model.half().cuda()  # fp16
tokenizer = AutoTokenizer.from_pretrained("THUDM/chatglm3-6b", trust_remote_code=True)

sents = ['对下面中文拼写纠错：\n少先队员因该为老人让坐。',
         '对下面中文拼写纠错：\n下个星期，我跟我朋唷打算去法国玩儿。']
for s in sents:
    response = model.chat(tokenizer, s, max_length=128, eos_token_id=tokenizer.eos_token_id)
    print(response)
```

output:
```shell
少先队员应该为老人让座。
下个星期，我跟我朋友打算去法国玩儿。
```


模型文件组成：
```
chatglm3-6b-csc-chinese-lora
    ├── adapter_config.json
    └── adapter_model.bin
```

#### 训练参数：

![loss](train_loss.png)

- num_epochs: 5
- per_device_train_batch_size: 6
- learning_rate: 2e-05
- best steps: 25100
- train_loss: 0.0834
- lr_scheduler_type: linear
- base model: THUDM/chatglm3-6b
- warmup_steps: 50
- "save_strategy": "steps"
- "save_steps": 500
- "save_total_limit": 10
- "bf16": false
- "fp16": true
- "optim": "adamw_torch"
- "ddp_find_unused_parameters": false
- "gradient_checkpointing": true
- max_seq_length: 512
- max_length: 512
- prompt_template_name: vicuna
- 6 * V100 32GB, training 48 hours

### 训练数据集
训练集包括以下数据：

- 中文拼写纠错数据集：https://huggingface.co/datasets/shibing624/CSC
- 中文语法纠错数据集：https://github.com/shibing624/pycorrector/tree/llm/examples/data/grammar
- 通用GPT4问答数据集：https://huggingface.co/datasets/shibing624/sharegpt_gpt4


如果需要训练GPT模型，请参考[https://github.com/shibing624/pycorrector](https://github.com/shibing624/pycorrector)



## Citation

```latex
@software{pycorrector,
  author = {Ming Xu},
  title = {pycorrector: Text Error Correction Tool},
  year = {2023},
  url = {https://github.com/shibing624/pycorrector},
}
```