TsitkoD/Qwen3-14B-Vedun-v5-bf16

TsitkoD/Qwen3-14B-Vedun-v5-bf16: Ancient Slavic Bukvitsa Interpretation

This model is a 14 billion parameter Qwen3-based language model, fine-tuned by TsitkoD using LoRA on a synthetic dataset (TsitkoD/vedun-lora-data). Its core function is to answer Russian-language questions by performing a structural analysis based on the Ancient Slavic Bukvitsa tradition.

Key Capabilities

• Bukvitsa-based Analysis: Deconstructs words into individual 'Bukvitsy', detailing each letter's name, numerical value, and associated imagery.
• Contextual Interpretation: Synthesizes the meanings of individual letters to form a comprehensive understanding of the word's sense.
• Lore Integration: Incorporates references to the Veles Book and Bukvitsa tradition to enrich its explanations.
• Multi-term Question Handling: For complex or 'everyday' questions (e.g., "how to forgive betrayal?"), the model identifies 2-4 key terms, analyzes each separately via Bukvitsa, and then synthesizes a cohesive answer.
• Stylized Output: Generates responses in a distinct aesthetic of "ancient Slavic Bukvitsa," providing a unique textual experience.

Training Details

The model underwent a three-stage LoRA fine-tuning process, progressively enhancing its ability to perform Bukvitsa analyses and handle multi-term questions. The bf16 variant represents the highest fidelity version, with q8 and q4 quantized versions available for memory efficiency, though q4 may occasionally lose accuracy in letter-level breakdowns.

Limitations

• The model's interpretations are a stylization and not academic linguistic truth.
• It will attempt Bukvitsa analysis even for questions outside its primary domain.
• Quantized versions, especially q4, may exhibit reduced accuracy in letter identification.