KLXTdwz/ChineseErrorCorrector4-4B

TEXT GENERATIONPricing:Input $0.4 / Cached $0.08 / Output $0.8Concurrent Unit Cost:1Model Size:4BQuant:BF16Context Size:32kTool Calling:SupportedPublished:Sep 2, 2026License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Cold

KLXTdwz/ChineseErrorCorrector4-4B is a 4 billion parameter Chinese Grammatical Error Correction (CGEC) and Chinese Spelling Check (CSC) model developed by KLXTdwz. Built on the CSRP three-stage training framework, it addresses over-correction bias by internalizing linguistic priors, distilling Chain-of-Thought reasoning, and using an Efficiency-Aware Reward system. This model achieves state-of-the-art performance on both NACGEC and CSCD benchmarks, making it highly effective for precise Chinese text correction.

Loading preview...

Overview

ChineseErrorCorrector4-4B (CSRP) is a 4 billion parameter model specifically designed for high-precision Chinese Grammatical Error Correction (CGEC) and Chinese Spelling Check (CSC). Developed by KLXTdwz, this model was presented in the paper "CSRP: Chain-of-Thought Reasoning for Chinese Text Correction via Reinforcement Learning with Efficiency-Aware Rewards," accepted as an Oral presentation at ACL 2026.

Key Capabilities and Innovations

This model is built upon a unique CSRP (CPT → SFT → RL) three-stage training framework, which effectively mitigates the common problem of over-correction bias in traditional LLM-based correction systems. The stages include:

  • Balanced Continued Pre-training (CPT): Internalizes linguistic priors using a large dataset with a mixture of general and correction-specific data.
  • Rationale-Augmented SFT: Distills Chain-of-Thought reasoning paths, enabling the model to diagnose error types before making corrections.
  • Efficiency-Aware Policy Alignment: Utilizes GRPO with a novel Efficiency-Aware Reward (EAR) to penalize unnecessary edits and promote surgical precision in corrections.

Performance Highlights

ChineseErrorCorrector4-4B has achieved state-of-the-art (SOTA) results on key benchmarks:

  • CGEC (NACGEC Benchmark): Achieved an $F_{0.5}$ score of 50.99, significantly surpassing previous leading models like ScholarGEC (14B) and CEC3 (4B).
  • CSC (CSCD Benchmark): Demonstrated strong performance with a Correction F1 score of 59.61, outperforming models including GPT-4 (Few-shot).

Use Cases

This model is ideal for applications requiring highly accurate and precise Chinese text correction, such as:

  • Automated proofreading and editing tools.
  • Educational platforms for Chinese language learners.
  • Content creation and publishing workflows to ensure grammatical correctness and spelling accuracy.

Finetunes
1 models