GRMR-V3-Q1.7B: Specialized Grammar Correction Model
GRMR-V3-Q1.7B, developed by qingy2024, is a 2 billion parameter Qwen3-based model fine-tuned from unsloth/Qwen3-1.7B-Base. Its primary purpose is grammar correction, addressing grammatical errors, punctuation, spelling, and overall text quality in English.
Key Capabilities
- Grammar Correction: Fixes a wide range of grammatical mistakes.
- Punctuation & Spelling: Corrects common punctuation and spelling errors.
- Text Quality Improvement: Enhances sentence structure and clarity.
- Specialized Chat Template: Uses
text_start/text_end for input and corrected_start/corrected_end for output, ensuring clear distinction.
Training Details
The model underwent full parameter fine-tuning on the qingy2024/grmr-v4-60k dataset, which comprises 60,000 grammar correction examples. Training utilized the Unsloth framework with a batch size of 8, 1 epoch, and a max sequence length of 16,384.
Recommended Usage
For optimal results, specific sampler settings are recommended: temperature = 0.7, frequency_penalty = 0.0, presence_penalty = 0.0, min_p = 0.01, top_p = 0.95, and top_k = 40.
Limitations
- May struggle with highly technical or domain-specific content.
- Performance can vary for non-standard English or text with multiple errors.
- Focuses on grammatical correctness, not stylistic improvements, and may occasionally introduce errors or alter meaning.