IvanCvetanovic/Slava-Qwen3-14B-Serbian

Slava: A Serbian Multi-Task LLM

Slava is a 14 billion parameter Qwen3-based model, developed by Ivan Cvetanović as part of a master's thesis at the University of Vienna. It is specifically full-fine-tuned for Serbian multi-task NLP, supporting both Cyrillic and Latin scripts. The model was trained for a single epoch on a mixture of 24,601 instruction-formatted examples across five generative tasks.

Key Capabilities

• Grammar Correction: Significantly improved performance over the base model, achieving an exact match score of 0.455.
• Serbian-to-English Translation: Maintains strong BLEU scores (61.75).
• Summarization: While automatic metrics show improvement, native-speaker studies indicated a preference for the base model's summaries due to Slava's more compressed output.
• Sentiment Analysis: Achieves 82.7% accuracy on ParlaSent and 76.8% on SerbMR datasets.
• Aspect-Based Sentiment Analysis (ABSA): Improved Aspect-F1 to 0.622.

Training and Evaluation

Slava was fine-tuned from OpenPipe/Qwen3-14B-Instruct using a single epoch and an effective batch size of 16. Training involved a diverse dataset including Serbian Wikipedia, OPUS-100, SentiComments.SR, and synthetic reviews generated by Qwen2.5-72B-Instruct. Evaluation showed substantial gains in specific Serbian NLP tasks, while general knowledge benchmarks remained consistent with the base model, indicating the fine-tuning's targeted impact.

Good For

• Research and non-commercial use in Serbian NLP.
• Instruction-style grammar correction, translation, summarization, and sentiment analysis in Serbian.

Limitations

• Non-commercial license (CC BY-NC-SA 4.0) due to training data sources.
• Summaries can be shorter and potentially less faithful than the base model.
• Not intended for high-stakes or safety-critical applications.
• Reasoning-heavy tasks (e.g., moral scenarios, mathematics) remain areas of weakness.