Qwen3-4B German Teacher: Specialized Grammar Instruction

This model, developed by Jaigouk Kim, is a 4 billion parameter Qwen3-based language model specifically fine-tuned for German language teaching, targeting A1-B1 CEFR levels. It leverages Supervised Fine-Tuning (SFT) with LoRA to achieve its specialized capabilities.

Key Capabilities

• Grammar Error Detection: Identifies grammatical mistakes in German sentences.
• Error Correction: Provides correct forms with clear, learner-friendly explanations.
• Grammar Judgment: Classifies sentence grammaticality (CoLA-style).
• Teaching Explanations: Offers clear explanations of German grammar rules, focusing on common pitfalls like haben/sein auxiliary verbs, article gender, case usage, and word order.

Performance Highlights

Evaluated on a custom German grammar benchmark, the model demonstrates strong performance:

• CoLA MCC: 0.721 for grammaticality judgment.
• GEC F1: 0.349 for grammar error correction.
• Generation Quality: 3.99/5.0 in human evaluation.

It significantly outperforms models like SmolLM3-German-V6, showing a +15.5% improvement in CoLA MCC and a +140.7% improvement in GEC F1, indicating superior accuracy in grammar tasks. The training data comprised ~9,000 examples, with a strong focus on grammar correction, judgment, and structured teaching content, informed by research on instruction-tuning and avoiding the "alignment tax" of preference optimization for accuracy-critical tasks.

Ideal Use Cases

• German Language Learning Apps: Integrating grammar checking and explanation features.
• Educational Tools: Assisting A1-B1 German learners with grammar practice and understanding.
• Automated Feedback Systems: Providing targeted feedback on written German assignments.