qwen3-4b-german-teacher-v1/README.md

---
license: apache-2.0
base_model: Qwen/Qwen3-4B
tags:
  - german
  - language-learning
  - grammar
  - education
  - qwen3
  - finetuned
  - sft
language:
  - de
  - en
pipeline_tag: text-generation
library_name: transformers
datasets:
  - custom
model-index:
  - name: qwen3-4b-german-teacher
    results:
      - task:
          type: text-generation
          name: German Grammar Teaching
        metrics:
          - type: cola_mcc
            value: 0.721
            name: CoLA MCC (Grammaticality Judgment)
          - type: gec_f1
            value: 0.349
            name: GEC Macro F1 (Error Correction)
          - type: generation_quality
            value: 3.99
            name: Generation Quality (1-5 scale)
          - type: overall_score
            value: 0.633
            name: Overall Score
---

# Qwen3-4B German Teacher

A finetuned [Qwen3-4B](https://huggingface.co/Qwen/Qwen3-4B) model specialized for German language teaching at A1-B1 CEFR levels. This model excels at:

- **Grammar Error Detection**: Identifying grammatical mistakes in German sentences
- **Error Correction**: Providing correct forms with clear explanations
- **Grammar Judgment**: Binary classification of sentence grammaticality (CoLA-style)
- **Teaching Explanations**: Clear, learner-friendly explanations of German grammar rules

## Model Details

| Property | Value |
|----------|-------|
| Base Model | Qwen/Qwen3-4B |
| Parameters | 4B |
| Training Method | SFT (Supervised Fine-Tuning) with LoRA |
| LoRA Rank | 32 |
| LoRA Alpha | 64 |
| Training Epochs | 2 |
| Learning Rate | 2e-4 |
| Context Length | 4096 tokens |

## Performance

Evaluated on a custom German grammar benchmark:

| Metric | Score | Description |
|--------|-------|-------------|
| **CoLA MCC** | 0.721 | Matthews Correlation Coefficient for grammaticality judgment |
| **GEC F1** | 0.349 | Macro F1 for grammar error correction |
| **Generation Quality** | 3.99/5.0 | Human-evaluated response quality |
| **Overall Score** | 0.633 | Weighted composite score |

### Benchmark Comparison

Comparison with SmolLM3-German-V6 (3B parameters, Q4 quantized):

| Metric | Qwen3 German Teacher (4B) | SmolLM3-German-V6 (3B) | Improvement |
|--------|---------------------------|------------------------|-------------|
| **CoLA MCC** | **0.721** | 0.624 | +15.5% |
| **GEC F1** | **0.349** | 0.145 | +140.7% |
| **Generation Quality** | **3.99** | 3.19 | +25.1% |
| **Overall Score** | **0.633** | 0.492 | +28.7% |

Key advantages of this model:
- **Superior grammaticality judgment**: 15.5% higher CoLA MCC score
- **Much better error correction**: 2.4x better GEC F1 score
- **Higher quality responses**: 0.8 points higher on 5-point scale

## Training Data

The model was trained on ~9,000 examples with the following composition:

| Category | Percentage | Purpose |
|----------|------------|---------|
| Grammar Correction | 35% | Error correction patterns |
| Grammar Judgment | 25% | CoLA-style "Is this correct?" examples |
| Structured Teaching | 20% | Verb conjugations, grammar explanations |
| General Conversation | 20% | Fluency preservation |

### Key Training Focus Areas

- **haben/sein auxiliary verbs**: Movement verbs require "sein" (e.g., "Ich bin gefahren" not "Ich habe gefahren")
- **Article gender** (der/die/das)
- **Case usage** (Nominativ, Akkusativ, Dativ, Genitiv)
- **Word order** in German sentences

## Usage

### With Transformers

```python
from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained("jaigouk/qwen3-4b-german-teacher")
tokenizer = AutoTokenizer.from_pretrained("jaigouk/qwen3-4b-german-teacher")

messages = [
    {"role": "system", "content": "Du bist ein freundlicher Deutschlehrer für A1-B1 Lernende."},
    {"role": "user", "content": "Is this German sentence correct? 'Ich habe nach Berlin gefahren.'"}
]

text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tokenizer(text, return_tensors="pt")
outputs = model.generate(**inputs, max_new_tokens=256)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))
```

### With Ollama (GGUF)

1. Download the GGUF file: `qwen3-german-teacher-Q4_K_M.gguf`

2. Create a Modelfile:
```dockerfile
FROM ./qwen3-german-teacher-Q4_K_M.gguf

TEMPLATE """{{- if .System }}<|im_start|>system
{{ .System }}<|im_end|>
{{ end }}<|im_start|>user
{{ .Prompt }}<|im_end|>
<|im_start|>assistant
<think>

</think>

{{ .Response }}<|im_end|>
"""

PARAMETER stop "<|im_end|>"
PARAMETER stop "<|endoftext|>"
PARAMETER temperature 0.7
PARAMETER top_p 0.9
PARAMETER num_ctx 4096

SYSTEM """Du bist ein freundlicher Deutschlehrer für A1-B1 Lernende. Erkläre Grammatikfehler klar und einfach. Beurteile grammatische Korrektheit sorgfältig."""
```

3. Create and run:
```bash
ollama create qwen3-german-teacher -f Modelfile
ollama run qwen3-german-teacher "Is 'Ich habe nach Berlin gefahren' correct?"
```

## Example Interactions

### Grammar Error Detection

**User**: Is this German sentence grammatically correct? "Ich habe nach Berlin gefahren."

**Model**: No, this sentence is incorrect. The verb "fahren" describes movement from one place to another, so it requires "sein" as the auxiliary verb, not "haben".

**Correct**: "Ich **bin** nach Berlin gefahren."

### Grammar Explanation

**User**: Explain when to use "sein" vs "haben" in the Perfekt tense.

**Model**: In German, most verbs use "haben" in the Perfekt tense, but verbs that describe:
1. **Movement from A to B**: gehen, fahren, fliegen, kommen, laufen
2. **Change of state**: aufwachen, einschlafen, sterben, werden

These verbs use "sein" instead. For example:
- Ich **bin** nach Hause gegangen. (movement)
- Er **ist** eingeschlafen. (change of state)

### Conjugation Help

**User**: Conjugate "sein" in present tense.

**Model**: Here is the conjugation of "sein" (to be) in present tense:

| Person | Conjugation |
|--------|-------------|
| ich | bin |
| du | bist |
| er/sie/es | ist |
| wir | sind |
| ihr | seid |
| sie/Sie | sind |

## Limitations

- Optimized for A1-B1 level German; may oversimplify advanced grammar concepts
- Focused on grammar teaching; not designed for creative writing or translation
- Best performance with clear, focused grammar questions

## Training Methodology

### Research-Informed Approach

This model was developed using insights from recent research on instruction-tuning and continual learning:

#### GRAPE Data Selection

Based on the [GRAPE paper (Feb 2025)](https://arxiv.org/abs/2502.04194):

> "Instructions paired with responses sampled from other LLMs are often **out of the distribution of the target model**, which can lead to diminishing returns and even hurt performance"

**Key insight**: Rather than using responses generated by larger models (GPT-4, Claude), we curated training examples that align with Qwen3's natural output distribution. This prevents the model from learning patterns it cannot reliably reproduce.

#### CoLA-Style Judgment Training

The breakthrough in V4 was adding 25% **explicit grammaticality judgment examples**. This teaches the model to reason about correctness rather than just generate correct text:

```
User: Ist der Satz grammatisch korrekt? "Der Frau ist schön."
Assistant: Nein, der Satz ist nicht korrekt.
Fehler: "Der" ist maskulin, aber "Frau" ist feminin.
Korrektur: "Die Frau ist schön."
```

#### Lessons from EWC/Fisher Information Research

From [arXiv:2502.11756](https://arxiv.org/html/2502.11756v1) on Fisher Information computation:
- **EXACT computation outperforms approximations**
- Minimum 500 samples for reliable Fisher estimation
- Device consistency (GPU-only) is critical

These insights from our earlier SmolLM3 experiments (V5/V6 with Elastic Weight Consolidation) informed our dataset composition decisions for Qwen3-V4.

## Training Details

### Multi-Stage Finetuning: Lessons Learned

This project was inspired by the 3-stage finetuning approach described in [MiroThinker (arXiv:2511.11793)](https://arxiv.org/abs/2511.11793):

1. **Stage 1: SFT** (Supervised Fine-Tuning)
2. **Stage 2: DPO** (Direct Preference Optimization)
3. **Stage 3: RL/GRPO** (Reinforcement Learning with Grammar Rewards)

Our training script design followed the MiroThinker paper's architecture, including:
- **Higher LoRA rank (r=32, alpha=64)** as recommended for multi-stage training stability
- **Full target modules** (q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj) for comprehensive adaptation

However, **this model uses only SFT** because our experiments with preference optimization showed a fundamental trade-off:

#### Why DPO/SimPO Failed for Our Use Case

| Model | CoLA MCC | GEC F1 | Generation | Overall | Status |
|-------|----------|--------|------------|---------|--------|
| **V6 SFT (Baseline)** | **0.624** | **0.191** | 3.29 | **0.516** | Reference |
| V8 (DPO) | 0.583 | 0.063 | 3.63 | 0.473 | -8% overall |
| V9 (SimPO) | 0.567 | 0.073 | **3.72** | 0.479 | -7% overall |

**Key findings**:
- DPO improved generation quality (+10%) but **destroyed GEC accuracy (-67%)**
- SimPO achieved best generation (3.72) but still regressed accuracy significantly
- This confirms the "alignment tax" documented in 2025 research: 76% of preference optimization causes regression on specific tasks

For grammar teaching, **accuracy is more important than fluency**, so we stayed with pure SFT. The V4 dataset composition (25% CoLA-style judgment examples) proved more effective than preference optimization for our metrics.

### SFT Training Configuration

Configuration based on [MiroThinker](https://arxiv.org/abs/2511.11793) recommendations for multi-stage training:

```
Base Model: Qwen/Qwen3-4B
LoRA Configuration:
  - Rank (r): 32
  - Alpha: 64
  - Target modules: q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj
  - Dropout: 0

Training Configuration:
  - Epochs: 2
  - Learning Rate: 2e-4 (linear decay)
  - Batch Size: 2 (effective: 8 with gradient accumulation)
  - Warmup: 10%
  - Precision: BF16
  - Optimizer: AdamW 8-bit
```

The SFT stage trains on ~9,000 curated examples covering grammar judgment, error correction, and teaching explanations.

### Deployment Pipeline

After SFT training, the model goes through:

#### 1. PEFT Merge

LoRA adapters are merged into the base model:

```python
from peft import PeftModel
from transformers import AutoModelForCausalLM

base_model = AutoModelForCausalLM.from_pretrained("Qwen/Qwen3-4B")
model = PeftModel.from_pretrained(base_model, "lora_adapters/")
model = model.merge_and_unload()
model.save_pretrained("merged_model/", safe_serialization=True)
```

#### 2. GGUF Quantization

For efficient local inference:

```bash
python convert_hf_to_gguf.py merged_model/ --outtype bf16 --outfile model-bf16.gguf
llama-quantize model-bf16.gguf model-Q4_K_M.gguf Q4_K_M
```

The Q4_K_M quantization reduces model size from ~8GB to 2.5GB while maintaining high quality.

### Technical Specifications

- **Framework**: Unsloth + Transformers + TRL + PEFT
- **Hardware**: NVIDIA RTX 4090 (24GB VRAM)
- **Training Time**: ~45 minutes for 2 epochs
- **Quantization**: GGUF Q4_K_M (2.5GB)

## Citation

If you use this model, please cite:

```bibtex
@misc{qwen3-4b-german-teacher-2025,
  author = {Jaigouk Kim},
  title = {Qwen3-4B German Teacher: A Finetuned Model for German Grammar Teaching},
  year = {2025},
  publisher = {HuggingFace},
  url = {https://huggingface.co/jaigouk/qwen3-4b-german-teacher}
}
```

## License

This model is released under the Apache 2.0 license, following the base Qwen3 model license.

## Acknowledgments

- [Qwen Team](https://github.com/QwenLM/Qwen3) for the excellent base model
- [Unsloth](https://github.com/unslothai/unsloth) for efficient finetuning tools