Add dataset card

README.md

@@ -0,0 +1,165 @@
+---
+license: cc-by-4.0
+language:
+- ru
+pretty_name: Russian NMO Medical MCQ
+task_categories:
+- question-answering
+- text-generation
+tags:
+- russian
+- medical
+- nmo
+- multiple-choice
+- supervised-fine-tuning
+configs:
+- config_name: mcq
+  data_files:
+  - split: train
+    path: data/mcq/train-*.parquet
+  - split: validation
+    path: data/mcq/validation-*.parquet
+  - split: test
+    path: data/mcq/test-*.parquet
+- config_name: sft_single
+  data_files:
+  - split: train
+    path: data/sft_single/train-*.parquet
+  - split: validation
+    path: data/sft_single/validation-*.parquet
+  - split: test
+    path: data/sft_single/test-*.parquet
+- config_name: sft_multi
+  data_files:
+  - split: train
+    path: data/sft_multi/train-*.parquet
+  - split: validation
+    path: data/sft_multi/validation-*.parquet
+  - split: test
+    path: data/sft_multi/test-*.parquet
+- config_name: sft_all
+  data_files:
+  - split: train
+    path: data/sft_all/train-*.parquet
+  - split: validation
+    path: data/sft_all/validation-*.parquet
+  - split: test
+    path: data/sft_all/test-*.parquet
+---
+
+# Russian NMO Medical MCQ
+
+Russian-language NMO medical test questions with single-correct and multiple-correct
+multiple-choice formats, prepared for LLM fine-tuning and evaluation.
+
+This dataset is intended for educational test-answering, supervised fine-tuning,
+retrieval/evaluation experiments, and benchmark construction. It is not clinical
+decision support.
+
+## Configs
+
+- `mcq`: canonical structured examples.
+- `sft_single`: chat/instruction rows for single-correct-choice questions.
+- `sft_multi`: chat/instruction rows for multiple-correct-choice questions.
+- `sft_all`: concatenation of `sft_single` and `sft_multi`.
+
+## Loading
+
+```python
+from datasets import load_dataset
+
+mcq = load_dataset("drkolesnikov/russian-nmo-medical-mcq", "mcq")
+sft_all = load_dataset("drkolesnikov/russian-nmo-medical-mcq", "sft_all")
+```
+
+For a local export:
+
+```python
+from datasets import load_dataset
+
+mcq = load_dataset(".", "mcq")
+sft_all = load_dataset(".", "sft_all")
+```
+
+## Schema
+
+`mcq` columns:
+
+```text
+id
+topic
+question
+choices
+correct_choice_indices
+correct_choice_texts
+choice_count
+correct_choice_count
+is_multi_answer
+task_type
+```
+
+SFT columns:
+
+```text
+id
+topic
+task_type
+correct_choice_indices
+messages
+prompt
+completion
+```
+
+`task_type` is one of:
+
+- `single_choice`
+- `multiple_correct`
+
+Answer indices are zero-based and preserve the original answer-choice order.
+
+## Prompt Format
+
+SFT prompts include the topic when available, then the question and zero-based
+answer choices. SFT completions contain only selected answer number(s) and answer
+text, with no explanations.
+
+Single-correct example:
+
+```text
+Ответ: 2. системный воспалительный ответ
+```
+
+Multiple-correct example:
+
+```text
+Ответы: 0, 2, 4. лихорадка; тахикардия; лейкоцитоз
+```
+
+## Splits
+
+Splits are deterministic and grouped by normalized question ID:
+
+- train: 90%
+- validation: 5%
+- test: 5%
+
+All variants derived from the same normalized question stay in the same split.
+
+## Counts
+
+| Config | Train | Validation | Test | Total |
+| --- | ---: | ---: | ---: | ---: |
+| `mcq` | 384,286 | 21,332 | 21,342 | 426,960 |
+| `sft_single` | 260,561 | 14,546 | 14,640 | 289,747 |
+| `sft_multi` | 123,725 | 6,786 | 6,702 | 137,213 |
+| `sft_all` | 384,286 | 21,332 | 21,342 | 426,960 |
+
+Excluded from all public configs:
+
+- conflicting answer variants: 449
+- answer sets attached to conflicting variants: 914
+- answer sets with markdown/source-page artifacts: 4
+
+## License
+
+Released under `cc-by-4.0`.