ChineseErrorCorrector3-4B

๐Ÿ‡จ๐Ÿ‡ณไธญๆ–‡ | English

ChineseErrorCorrector3-4B is part of a comprehensive platform for Chinese text correction, integrating academic research, model training, evaluation, and inference. It covers two core directions: Spelling Correction (CSC) and Grammatical Error Correction (CGEC).

The methodology behind this line of models is presented in the paper CSRP: Chain-of-Thought Reasoning for Chinese Text Correction via Reinforcement Learning with Efficiency-Aware Rewards.

ๆจกๅž‹ๆ่ฟฐ

ไธ€ไธช้ขๅ‘ไธญๆ–‡ๆ–‡ๆœฌ็บ ้”™ไปปๅŠก็š„็ปผๅˆๅนณๅฐ๏ผŒ้›†ๅญฆๆœฏ็ ”็ฉถใ€ๆจกๅž‹่ฎญ็ปƒใ€ๆจกๅž‹่ฏ„ๆต‹ๅ’ŒๆŽจ็†้ƒจ็ฝฒไบŽไธ€ไฝ“๏ผŒ่ฆ†็›–ๆ‹ผๅ†™็บ ้”™ไธŽ่ฏญๆณ•็บ ้”™ไธคไธชๆ ธๅฟƒๆ–นๅ‘ใ€‚

  • twnlp/ChineseErrorCorrector3-4B: ไฝฟ็”จ200ไธ‡็บ ้”™ๆ•ฐๆฎ่ฟ›่กŒๅ…จ้‡่ฎญ็ปƒ๏ผŒ้€‚็”จไบŽ่ฏญๆณ•็บ ้”™ๅ’Œๆ‹ผๅ†™็บ ้”™๏ผŒๆ•ˆๆžœๆœ€ๅฅฝ๏ผŒๆŽจ่ไฝฟ็”จใ€‚

ๆจกๅž‹่ฏ„ๆต‹๏ผˆNaCGEC Data๏ผ‰

Model Name Base Model Avg SIGHAN-2015 EC-LAW MCSC GPU QPS
ChatGLM3-6B-CSC THUDM/chatglm3-6b 0.4538 0.6572 0.4369 0.2672 GPU 3
Qwen2.5-1.5B-CTC Qwen/Qwen2.5-1.5B-Instruct 0.6802 0.3032 0.7846 0.9529 GPU 6
Qwen2.5-7B-CTC Qwen/Qwen2.5-7B-Instruct 0.8225 0.4917 0.9798 0.9959 GPU 3
Qwen3-4B-CTC(Our) Qwen/Qwen3-4B 0.8521 0.6340 0.9360 0.9864 GPU 5

Sample Usage

You can use the model with the transformers library as follows:

from transformers import AutoModelForCausalLM, AutoTokenizer

model_name = "twnlp/ChineseErrorCorrector3-4B"

model = AutoModelForCausalLM.from_pretrained(
    model_name,
    torch_dtype="auto",
    device_map="auto"
)
tokenizer = AutoTokenizer.from_pretrained(model_name)

prompt = "ไฝ ๆ˜ฏไธ€ไธชๆ–‡ๆœฌ็บ ้”™ไธ“ๅฎถ๏ผŒ็บ ๆญฃ่พ“ๅ…ฅๅฅๅญไธญ็š„่ฏญๆณ•้”™่ฏฏ๏ผŒๅนถ่พ“ๅ‡บๆญฃ็กฎ็š„ๅฅๅญ๏ผŒ่พ“ๅ…ฅๅฅๅญไธบ๏ผš"
text_input = "ๅฏนๅพ…ๆฏไธ€้กนๅทฅไฝœ้ƒฝ่ฆไธ€ไธไธๅคŸใ€‚"
messages = [
    {"role": "user", "content": prompt + text_input}
]
text = tokenizer.apply_chat_template(
        messages,
        tokenize=False,
        add_generation_prompt=True,
        enable_thinking=False # Switches between thinking and non-thinking modes. Default is True.
    )
model_inputs = tokenizer([text], return_tensors="pt").to(model.device)

generated_ids = model.generate(
    **model_inputs,
    max_new_tokens=512
)
generated_ids = [
    output_ids[len(input_ids):] for input_ids, output_ids in zip(model_inputs.input_ids, generated_ids)
]

response = tokenizer.batch_decode(generated_ids, skip_special_tokens=True)[0]
print(response)
# Output: ๅฏนๅพ…ๆฏไธ€้กนๅทฅไฝœ้ƒฝ่ฆไธ€ไธไธ่‹Ÿใ€‚

Citation

If you find this work helpful, please cite:

@misc{tian2026csrpchainofthoughtreasoningchinese,
      title={CSRP: Chain-of-Thought Reasoning for Chinese Text Correction via Reinforcement Learning with Efficiency-Aware Rewards}, 
      author={Wei Tian and Yuhao Zhou and Man Lan},
      year={2026},
      eprint={2606.00020},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2606.00020}, 
}

@misc{tian2025chineseerrorcorrector34bstateoftheartchinesespelling,
      title={ChineseErrorCorrector3-4B: State-of-the-Art Chinese Spelling and Grammar Corrector},
      author={Wei Tian and YuhaoZhou},
      year={2025},
      eprint={2511.17562},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2511.17562},
}
Downloads last month
203
Safetensors
Model size
4B params
Tensor type
BF16
ยท
Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐Ÿ™‹ 1 Ask for provider support

Model tree for twnlp/ChineseErrorCorrector3-4B

Finetuned
Qwen/Qwen3-4B
Finetuned
(1,285)
this model
Quantizations
2 models

Space using twnlp/ChineseErrorCorrector3-4B 1

Papers for twnlp/ChineseErrorCorrector3-4B