Nubinu/Qwen3.5-4B-MiniFantasy
Nubinu/Qwen3.5-4B-MiniFantasy is a 4.5 billion parameter language model, fine-tuned from MuXodious/Qwen3.5-4B-SOMPOA-heresy-v2 using 4-bit LoRA. This model specializes in generating multi-turn narrative pacing and emotional resonance, particularly for character-driven interactions. It is optimized for roleplay scenarios, especially within platforms like SillyTavern, by adhering to specific character card formats. The model's training focused on creating dynamic and engaging character interactions.
Loading preview...
Model Overview
Nubinu/Qwen3.5-4B-MiniFantasy is a 4.5 billion parameter language model, fine-tuned from the MuXodious/Qwen3.5-4B-SOMPOA-heresy-v2 base model. This iteration utilizes a 4-bit LoRA (Low-Rank Adaptation) fine-tuning approach, specifically designed to enhance narrative generation capabilities.
Key Capabilities
- Narrative Pacing: Excels at maintaining coherent and engaging multi-turn narrative flows.
- Emotional Resonance: Trained to generate responses with appropriate emotional depth and character dynamics.
- Character Adherence: Highly effective at adhering to detailed character profiles when provided in a specific Markdown format.
- Roleplay Optimization: Fine-tuned on a custom synthetic dataset tailored for interactive roleplay scenarios.
Training Details
The model underwent 3 epochs of LoRA fine-tuning with a rank of 16 and alpha of 16, targeting key attention and feed-forward modules (q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj). It used a paged_adamw_8bit optimizer with a learning rate of 5e-5 and a cosine LR scheduler. The training dataset was custom-curated, focusing on synthetic data for narrative pacing and character interaction.
Recommended Use Cases
- SillyTavern and Roleplay: Ideal for generating dynamic and character-consistent responses in roleplaying environments. Users are encouraged to follow the recommended character card format for optimal performance.
- Interactive Storytelling: Suitable for applications requiring models to maintain complex character identities and narrative arcs over extended conversations.