Danetki Qwen3-0.6B: Specialized for Russian Lateral Thinking Puzzles
This model, developed by lifeart, is a Qwen3-0.6B base model fine-tuned with QLoRA specifically for the Russian game "Данетки" (lateral thinking puzzles). Its primary function is to act as the puzzle host, responding to player questions with a single word: Да (Yes), Нет (No), or Не важно (Irrelevant).
Key Capabilities & Features
- Specialized Response Format: Designed to output only one of three specific Russian words, making it highly focused for its task.
- Client-Side Deployment: Available for in-browser inference via WebGPU, requiring no server.
- Robust Training: Fine-tuned on approximately 183K training examples, including scraped Q&A pairs from danetka.ru and reformatted DaNetQA data.
- Performance: Achieves 76% accuracy and 74.6% Macro F1 score on a test set of 190 questions, with a slight bias towards "Да" responses.
- Language: Fully supports Russian language interactions.
Ideal Use Cases
- Interactive "Данетки" Game Host: Perfect for applications or platforms that simulate the "Данетки" game experience.
- Educational Tools: Can be integrated into language learning or puzzle-solving applications for Russian speakers.
- Research in Specialized LLMs: Demonstrates effective fine-tuning for highly specific, constrained natural language tasks.