Views
No views yet
Note: This is a research model. It does not constitute legal advice.
| Component | Configuration |
|---|---|
| Architecture | Llama 3.2 3B (4-bit quantized, MLX) |
| Fine-tuning | LoRA (rank 16, 16 layers), merged into base |
| Training Data | 10,000 German legal Q&A pairs |
| Framework | MLX (Apple Silicon) |
| Training | 500 steps, loss 1.874 |
| Format | MLX safetensors (4-bit) |
| Size | 1.87 GB |
1from mlx_lm import load, generate
2
3model, tokenizer = load("jonprompter/german-legal-qa-3b")
4
5messages = [
6 {"role": "system", "content": "Du bist ein deutschsprachiger Rechtsassistent. Dies ist keine Rechtsberatung."},
7 {"role": "user", "content": "Was besagt §823 BGB?"}
8]
9
10prompt = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
11response = generate(model, tokenizer, prompt=prompt, max_tokens=256)
12print(response)1pip install mlx-lm
2mlx_lm generate --model jonprompter/german-legal-qa-3b --prompt "Was besagt §823 BGB?"| Metric | Score |
|---|---|
| Token F1 | 0.218 |
| BERTScore F1 | 0.652 |
| Citation Accuracy | 0.059 |
1@software{german_legal_qa_2026,
2 title={German Legal QA Assistant},
3 author={jonprompter},
4 year={2026},
5 url={https://github.com/myaxyo/german-legal-qa}
6}