Qwen Typo Fixer - CoreML Standalone

Contents

• CoreML Models (3 components):
  • qwen-typo-fixer_embeddings.mlpackage - Token embeddings
  • qwen-typo-fixer_FFN_PF_lut4_chunk_01of01.mlpackage - FFN with prefill/infer functions
  • qwen-typo-fixer_lm_head_lut6.mlpackage - Language model head
• Tokenizer Files:
  • tokenizer.json - Main tokenizer configuration
  • tokenizer_config.json - Tokenizer metadata
  • vocab.json - Vocabulary mapping
  • merges.txt - BPE merge rules
  • config.json - Model configuration
• Python Implementation:
  • typo_fixer_complete.py - Complete working implementation

Usage

Prerequisites

Run Typo Correction

Example Output

Original:  'I beleive this is teh correct answr.'
Corrected: 'I believe this is the correct answer.'
Status:    ✅ Typos likely fixed!

Model Architecture

1. Embeddings: Converts token IDs to hidden representations
2. FFN+Prefill: Dual-function model with prefill and infer functions for efficient generation
3. LM Head: Final linear layer producing vocabulary logits

Performance

• Model Size: 596M parameters (Qwen3-0.6B based)
• Accuracy: 88.5% sentence accuracy on typo correction
• Inference Speed: ~300ms per correction
• Memory: Optimized for Apple Neural Engine acceleration

Fine-tuned Model Details

• Base Model: Qwen/Qwen3-0.6B
• Fine-tuned For: Typo correction with 88.5% accuracy
• Training Data: Enhanced typo correction dataset
• Original HuggingFace Model: mazhewitt/qwen-typo-fixer