icecubetr/GRMR-V3-G4B

Hugging Face
VISIONConcurrency Cost:1Model Size:4.3BQuant:BF16Ctx Length:32kPublished:May 4, 2026License:apache-2.0Architecture:Transformer Open Weights Warm

GRMR-V3-G4B by qingy2024 is a 4.3 billion parameter Gemma 3 model, fine-tuned from unsloth/gemma-3-4b-pt, specifically optimized for English grammar correction tasks. It excels at fixing grammatical errors, punctuation, and spelling, utilizing a specialized chat template for structured input and output. The model has a context length of 32768 tokens and is designed to improve text quality by addressing various language issues.

Loading preview...

GRMR-V3-G4B: Grammar Correction Model

GRMR-V3-G4B is a 4.3 billion parameter Gemma 3 model developed by qingy2024, fine-tuned from unsloth/gemma-3-4b-pt. Its primary purpose is grammar correction, making it a specialized tool for improving text quality.

Key Capabilities

  • Grammar Error Correction: Identifies and fixes grammatical mistakes.
  • Punctuation Correction: Addresses incorrect or missing punctuation.
  • Spelling Correction: Rectifies spelling errors.
  • Sentence Structure Improvement: Enhances clarity and structure of sentences.

Training and Features

The model was fine-tuned using full parameter fine-tuning on the qingy2024/grmr-v4-60k dataset, which contains 60,000 examples of original and corrected text. It utilizes a custom chat template with <|text_start|>/<|text_end|> for user inputs and <|corrected_start|>/<|corrected_end|> for model outputs, ensuring clear distinction between original and corrected content. Optimal performance is achieved with specific sampler settings, including a temperature of 0.7.

Limitations

While effective for general grammar correction, the model may face challenges with highly technical or domain-specific content, struggle with context-dependent rules, and its performance can vary for non-standard English or text with numerous errors. It focuses on correctness rather than stylistic enhancements.