mazhewitt/qwen-typo-fixer

Qwen Enhanced Typo Fixer

A fine-tuned Qwen model for typo correction using advanced error patterns and multi-domain training data.

Model Description

This model is a fine-tuned version of Qwen/Qwen2-0.5B for typo correction. It was trained on an enhanced dataset featuring:

• 80,677 training examples with realistic error patterns
• Multi-domain coverage: conversational, professional, educational, creative, instructional, general
• Advanced error types: keyboard errors, phonetic confusions, contextual mistakes, punctuation variations
• Balanced punctuation: 50/50 split between sentences with/without ending punctuation

Training Details

• Base Model: Qwen/Qwen2-0.5B
• Training Hardware: Dual RTX5090 (48GB total VRAM)
• Dataset Size: 80,677 examples
• Epochs: 3
• Batch Size: 32 (16 per GPU × 2 GPUs)
• Learning Rate: 5e-5

Usage

from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained("mazhewitt/qwen-typo-fixer", trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained("mazhewitt/qwen-typo-fixer", trust_remote_code=True)

# Example usage
text_with_typos = "I beleive this is teh correct answr."
prompt = f"<|im_start|>user\nCorrect the typos in this text: {text_with_typos}<|im_end|>\n<|im_start|>assistant\n"

inputs = tokenizer(prompt, return_tensors="pt")
outputs = model.generate(**inputs, max_new_tokens=50, temperature=0.1)
correction = tokenizer.decode(outputs[0], skip_special_tokens=True)
print(correction)

Training Configuration

The model was trained with the following key parameters:

• Learning rate: 5e-5
• Batch size: 32 (16 per GPU × 2 GPUs)
• Gradient accumulation: 2 steps
• Weight decay: 0.01
• Warmup ratio: 0.1

Dataset Features

Error Pattern Distribution

• Spelling errors: 60%+ coverage
• Keyboard errors: 30%+ coverage
• Phonetic errors: 8%+ coverage
• Grammar errors: 2%+ coverage
• Punctuation errors: <1% coverage

Domain Distribution

• Educational: 15,837 (19.6%)
• Instructional: 15,469 (19.2%)
• Creative: 15,254 (18.9%)
• Professional: 14,582 (18.1%)
• Conversational: 12,875 (16.0%)
• General: 6,660 (8.3%)

Complexity Distribution

• Simple: 39,159 (48.5%)
• Medium: 28,959 (35.9%)
• Complex: 12,559 (15.6%)

Evaluation

The model achieves strong performance on typo correction tasks, with particular strength in:

• Single and multi-word typos
• Contextual corrections
• Maintaining original meaning and style
• Handling various text domains

Limitations

• Optimized for English text
• Best performance on sentences under 150 characters
• May struggle with highly technical or domain-specific terminology
• Designed for typo correction, not general text improvement

Citation

If you use this model, please cite:

@misc{qwen-enhanced-typo-fixer,
  title={Qwen Enhanced Typo Fixer},
  author={mazhewitt},
  year={2025},
  url={https://huggingface.co/mazhewitt/qwen-typo-fixer}
}

Training Data Details

• Average Difficulty Score: 35.9
• Average Errors per Example: 1.8
• Top Error Types:
  • spelling: 70,135 occurrences
  • keyboard: 35,935 occurrences
  • phonetic: 9,875 occurrences
  • grammar: 1,833 occurrences
  • punctuation: 79 occurrences