Add dataset card

README.md

@@ -1,26 +1,52 @@
 ---
-dataset_info:
-  features:
-  - name: messages
-    list:
-    - name: content
-      dtype: string
-    - name: role
-      dtype: string
-  splits:
-  - name: train
-    num_bytes: 89423810
-    num_examples: 35538
-  - name: validation
-    num_bytes: 12458241
-    num_examples: 5000
-  download_size: 51303492
-  dataset_size: 101882051
-configs:
-- config_name: default
-  data_files:
-  - split: train
-    path: data/train-*
-  - split: validation
-    path: data/validation-*
+language:
+- de
+license: other
+task_categories:
+- text-generation
+size_categories:
+- 10K<n<100K
+tags:
+- german
+- deutsch
+- sft
+- instruction-tuning
+- conversational
 ---
+
+# german-sft-mix
+
+A German supervised fine-tuning (SFT) dataset of multi-domain instruction and chat
+conversations. Used to fine-tune
+[Mario12355/nanochat-de-537m](https://huggingface.co/Mario12355/nanochat-de-537m).
+
+## Format
+
+A single column `messages`: a list of chat turns, each
+`{"role": "user" | "assistant", "content": str}`. Conversations start with `user`
+and (almost always) end with `assistant`.
+
+## Splits
+
+| Split | Conversations |
+|---|---|
+| train | 35,538 |
+| validation | 5,000 |
+
+Turns per conversation: 2 (median), up to 16 — about 66% single-turn, 34% multi-turn.
+
+## Usage
+
+```python
+from datasets import load_dataset
+
+ds = load_dataset("Mario12355/german-sft-mix")
+print(ds["train"][0]["messages"])
+```
+
+## Notes
+
+- German-language instruction/chat data (partly from translated sources).
+- 166 training conversations end on a `user` turn (no final assistant reply); filter
+  these if your trainer requires assistant-terminated dialogues.
+- Derived from public instruction data — verify upstream licenses before commercial use.