HELPER: Russian corpus for psychological text analysis and fine-tuning Qwen for local emotion detection
Researchers from HSE University and VK created HELPER, a corpus of psychologically charged Russian texts with local emotion and psychological characteristic markup. They fine-tuned Qwen-3 models (8B and 14B) on this corpus to demonstrate improved local emotion analysis.
The team collected a corpus of psychologically loaded Russian texts from psychological consultation platform B17.ru, Yandex Reviews, and VK posts. They developed a custom markup scheme based on interviews with practicing psychologists, covering 14 emotions, psychological characteristics, therapist response evaluation, and client state assessment. To handle the complex multi-criteria markup, they built the HELPER annotation tool using .NET 10, Blazor, PostgreSQL, and Keycloak. They fine-tuned Qwen-3 8B and Qwen-3 14B models via LoRA on the corpus to demonstrate local emotion and psycholinguistic analysis capability.
- Сокращения
- LLM = Large Language Model — Большая языковая модель
- ML = Machine Learning — Машинное обучение
- UI = User Interface — Пользовательский интерфейс
- URL = Uniform Resource Locator — Унифицированный указатель ресурса
Source: Habr — хаб ML —
original