German Legal QA 3B: Domain-Adapted LLM for German Law

A LoRA-adapted Llama 3.2 3B model fine-tuned on German legal question-answering data. Designed for use with a RAG (Retrieval-Augmented Generation) pipeline that grounds answers in actual statutory text.

This model answers questions about German civil, criminal, and procedural law with paragraph citations (§-references).

Note: This model is for educational and research purposes only. It does not constitute legal advice.

Supported Legal Domains

BGB (Civil Code) · StGB (Criminal Code) · ZPO (Civil Procedure) · GG (Basic Law) · HGB (Commercial Code) · ArbGG (Labor Courts)

Model Details

The adapter was trained using Apple MLX on an M1 Pro MacBook, demonstrating efficient legal domain adaptation on consumer hardware.

Training Data

Trained on GerLayQA (EACL 2024) — real legal questions from German citizens paired with expert lawyer answers.

RAG Pipeline (Recommended)

This adapter works best combined with retrieval over German federal statutes. The full system indexes 103,000+ legal text chunks from 6,111 German federal laws.

Query → Embed → FAISS Search → Retrieve § Text → Generate Answer

With RAG, the model produces factually grounded answers citing the correct statutory provisions. Without RAG, the model may hallucinate legal content.

Usage

With MLX (Apple Silicon)

from mlx_lm import load, generate

model, tokenizer = load(
    "mlx-community/Llama-3.2-3B-Instruct-4bit",
    adapter_path="jonprompter/german-legal-qa-3b-lora"
)

prompt = """<|begin_of_text|><|start_header_id|>system<|end_header_id|>

Du bist ein deutschsprachiger Rechtsassistent. Beantworte rechtliche Fragen präzise und verständlich. Zitiere relevante Paragraphen wenn möglich. Dies ist keine Rechtsberatung.<|eot_id|><|start_header_id|>user<|end_header_id|>

Was besagt §823 BGB?<|eot_id|><|start_header_id|>assistant<|end_header_id|>

"""

response = generate(model, tokenizer, prompt=prompt, max_tokens=256)
print(response)

With Transformers (PyTorch)

from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel

base_model = AutoModelForCausalLM.from_pretrained("unsloth/Llama-3.2-3B-Instruct")
model = PeftModel.from_pretrained(base_model, "jonprompter/german-legal-qa-3b-lora")
tokenizer = AutoTokenizer.from_pretrained("unsloth/Llama-3.2-3B-Instruct")

Evaluation

Evaluated on held-out GerLayQA test set with RAG retrieval (50 samples):

Scores reflect comparison against verbose forum-style lawyer answers. The system provides concise statutory answers — different style, same meaning (shown by BERTScore).

Example Output

Q: Was besagt §823 BGB?

A: § 823 BGB besagt, dass wer vorsätzlich oder fahrlässig das Leben, den Körper, die Gesundheit, die Freiheit, das Eigentum oder ein sonstiges Recht eines anderen widerrechtlich verletzt, dem anderen zum Ersatz des daraus entstehenden Schadens verpflichtet ist.

Limitations

• Requires RAG retrieval for factually accurate answers
• Trained on German law only — not applicable to other jurisdictions
• Should not be used as a substitute for professional legal advice
• May produce hallucinations without retrieval context
• Early-stage model (500 training steps) — answer quality improves with further training

Full System

The complete RAG pipeline (data acquisition, indexing, training, evaluation, demo) is available at:

github.com/myaxyo/german-legal-qa

Features:

• Data pipeline: 6,111 German federal statutes indexed
• FAISS vector search: 103K chunks, sub-second retrieval
• Hybrid retrieval: keyword + semantic search
• Gradio demo: Interactive chatbot with source citations
• MLX training: Fine-tune on Apple Silicon in 30 minutes

Citation

@software{german_legal_qa_2026,
  title={German Legal QA Assistant},
  author={jonprompter},
  year={2026},
  url={https://github.com/myaxyo/german-legal-qa}
}