twnlp/ChineseErrorCorrector3-4B

Hugging Face
TEXT GENERATIONConcurrency Cost:1Model Size:4BQuant:BF16Ctx Length:32kTool Calling:SupportedPublished:Jun 20, 2025License:apache-2.0Architecture:Transformer0.0K Open Weights Warm

The twnlp/ChineseErrorCorrector3-4B model, developed by TW-NLP, is a 4-billion parameter model based on the Qwen3-4B architecture. It is specifically designed for comprehensive Chinese text correction, excelling in both spelling correction (CSC) and grammatical error correction (CGEC). Trained on 2 million correction data points, this model integrates academic research, training, evaluation, and inference for robust performance in identifying and rectifying errors in Chinese text.

Loading preview...

ChineseErrorCorrector3-4B Overview

ChineseErrorCorrector3-4B is a specialized model developed by TW-NLP, built upon the Qwen3-4B base, and is part of a comprehensive platform for Chinese text correction. This model is distinguished by its focus on two core areas: Spelling Correction (CSC) and Grammatical Error Correction (CGEC). It has been extensively trained on 2 million correction data points, making it highly effective for identifying and rectifying errors in Chinese text.

Key Capabilities

  • Comprehensive Chinese Text Correction: Addresses both spelling and grammatical errors.
  • High Performance: Achieves an average score of 0.8521 on the NaCGEC Data benchmark, outperforming other models like ChatGLM3-6B-CSC and Qwen2.5-7B-CTC in overall correction tasks.
  • Research-Backed Methodology: The model's development is informed by the paper "CSRP: Chain-of-Thought Reasoning for Chinese Text Correction via Reinforcement Learning with Efficiency-Aware Rewards" (arXiv:2606.00020).

Good For

  • Automated Chinese Text Proofreading: Ideal for applications requiring automatic detection and correction of errors in Chinese written content.
  • Academic and Professional Writing Tools: Can be integrated into tools for improving the quality of Chinese documents.
  • Natural Language Processing (NLP) Research: Serves as a strong baseline or component for further research in Chinese error correction.