Izbebe/med-assist-ru-sft

Dataset Card for MedAssist Russian Medical QA SFT

Dataset Structure

The dataset contains patient complaints in Russian (desc) along with the target medical specialty (spec10).

Fields:

• desc (string): The natural language text of the patient complaint or question.
• spec10 (string): Target medical specialty (e.g., Терапевт, Невролог, Хирург, Дерматолог, etc.).

Generation and Annotation Methodology

1. Source Data: Real medical forum questions from blinoff/medical_qa_ru_data.
2. Synthetic Data Generation (Self-Instruct): Diverse synthetic medical complaints generated using LLM prompts covering edge cases, typos, and varied sentence lengths.
3. Filtering & Deduplication: Cleaned for length thresholds (30–2000 chars) and deduplicated via Exact Match, MinHash, and Semantic Vector Similarity (cointegrated/rubert-tiny2).
4. Hybrid Annotation: Filtered using Confidence Score thresholds (Top1 - Top2 margin) with human-in-the-loop fallback simulation.

Class Distribution

• Total Records: 1,525
• Categories: 12 Medical Specialties (Терапевт, Гинеколог, Гастроэнтеролог, Невролог, Дерматолог, Офтальмолог, Отоларинголог, Хирург, Травматолог, Стоматолог, Проктолог, Уролог).

Example Entry

{
  "desc": "Уже 2 день мутная моча и жжение после туалета К какому врачу записаться?",
  "spec10": "Уролог"
}