GRMR-V3-G4B: Specialized Grammar Correction Model

GRMR-V3-G4B is a Gemma 3 4B model developed by qingy2024, meticulously fine-tuned for grammar correction in English text. It leverages full parameter fine-tuning on the extensive qingy2024/grmr-v4-60k dataset, which comprises 60,000 examples of original and corrected sentences.

Key Capabilities

• Grammar Error Correction: Identifies and rectifies grammatical mistakes.
• Punctuation and Spelling: Corrects punctuation and spelling errors.
• Sentence Structure Improvement: Enhances clarity and structure of sentences.
• Custom Chat Template: Utilizes a unique chat template with <|text_start|> and <|corrected_start|> tags to clearly delineate input and output.

Optimal Usage

For best performance, specific sampler settings are recommended, including temperature = 0.7, min_p = 0.01, top_p = 0.95, and top_k = 40. The model was trained with a maximum sequence length of 16,384, enabling it to handle longer texts. While highly effective for general grammar correction, it may face limitations with highly technical content or non-standard English, and might occasionally introduce subtle changes in meaning.