NPC Voice Model: soup-0.6
The walter-bd/npc-voice-soup06 is the best-performing model in the NPC voice series, built upon the Qwen3-0.6B architecture. This 0.8 billion parameter model is uniquely designed to transform plain factual sentences into character-specific dialogue, guided by six distinct persona parameters: TONE, STYLE, HUMOR, RELATION, and ROLE.
Key Capabilities & Features
- Persona-Conditioned Dialogue Generation: Rewrites factual input based on specified character attributes, enabling dynamic and consistent NPC voices.
- Hybrid Fine-tuning Approach: Developed as a "model soup," it's a weight-averaged merge (60% v5-SFT and 40% v5-DPO) to combine the strengths of both methods.
- Improved Reliability: This merging strategy significantly reduces verbatim copies and quote-wrapping failures from SFT, while mitigating over-correction and rare relation type forgetting seen in DPO.
- Configurable Parameters: Allows for fixed NPC characteristics (Tone, Style, Humor, Role) and dynamic, runtime-changeable parameters (Relation) to adapt to game state.
- Performance Metrics: Achieves a 61.5% pass rate, 28.5% hallucination failure, and 1.42 fact preservation score, outperforming its SFT and DPO predecessors.
Ideal Use Cases
- Dynamic NPC Dialogue: Generating contextually appropriate and character-consistent responses for non-player characters in video games.
- Interactive Storytelling: Creating varied narrative voices based on character profiles.
- Role-playing Applications: Providing distinct voices for different personas in text-based role-playing scenarios.