--- language: - de license: apache-2.0 tags: - austrian-german - instruction-tuning - sft - chat - austria pretty_name: Austrian German Instruction Dataset size_categories: - n<1K task_categories: - text-generation - question-answering extra_gated_prompt: >- This dataset is in validation phase. Access is granted to verified researchers and organizations. Please describe your intended use case. extra_gated_fields: Full name: text Organization: text Intended use: text I agree to use this data for research purposes only: type: checkbox --- # AT-Instruct: Austrian German Instructions 500 instruction-response pairs written in Austrian German. Not translated from English or Bundesdeutsch — written from scratch with Austrian vocabulary, institutions, and perspective. ## Why this exists Every German instruction dataset I found was either translated from English (losing all cultural context) or written in Bundesdeutsch. If you fine-tune on those, your model will tell users to go to the "Bürgeramt" — which doesn't exist in Austria. Here it's the Magistrat. So I wrote these manually. Erdapfel, not Kartoffel. ABGB, not BGB. Obers, not Sahne. ## What's in it | Category | Count | |----------|-------| | Wortschatz (vocab) | 108 | | Alltag (daily life) | 62 | | Kochen (cooking) | 53 | | Geographie | 45 | | Behördenweg (bureaucracy) | 40 | | Verkehr (transport) | 40 | | Kultur | 40 | | Gesundheit | 40 | | Bildung (education) | 36 | | Natur & Freizeit | 36 | ## Format ```json { "instruction": "Erklaere mir, wie ich mich in Wien anmelden kann.", "response": "Du gehst zum Magistratischen Bezirksamt deines Bezirks...", "category": "behoerdenweg", "register": "formal_austrian", "source": "hand_written" } ``` ## Usage ```python from datasets import load_dataset ds = load_dataset("Laborator/austrian-german-instructions") ``` ## Files - `data/at_instruct_v1.jsonl` - `data/at_instruct_v1.parquet` ## Known limitations 500 pairs is small for serious fine-tuning. I'm working on v2 with 2000+. But it's enough to see if an Austrian German adapter actually helps — and it does. ## License Apache 2.0