MuXodious/Qwen3.5-4B-MiniFantasy-MTP

MuXodious/Qwen3.5-4B-MiniFantasy-MTP Overview

This model is a 4.5 billion parameter Qwen3.5-based language model, fine-tuned by MuXodious. It is a 4-bit LoRA fine-tune of the MuXodious/Qwen3.5-4B-SOMPOA-heresy-v2 model, with Multi Token Prediction (MTP) weights restored from the base model. The fine-tuning focused on generating multi-turn narrative pacing, emotional resonance, and specific character dynamics.

Key Capabilities & Features

• Narrative Pacing: Optimized for fluid and engaging multi-turn narrative flows.
• Emotional Resonance: Designed to produce responses with nuanced emotional depth.
• Structured Character Adherence: Trained to follow a specific Markdown-based character card format for consistent personality and lore.
• LoRA Fine-tuning: Utilizes LoRA with a rank of 16 and alpha of 16, targeting key projection modules (q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj).
• Training Dataset: Fine-tuned on a custom, curated synthetic dataset.

Ideal Use Cases

• Creative Writing: Excellent for generating detailed stories, character interactions, and narrative arcs.
• Roleplay Applications: Specifically designed for platforms like SillyTavern, where consistent character portrayal and dynamic storytelling are crucial.
• Character Development: Useful for maintaining complex character identities and backstories through structured input.

Important Considerations

For optimal performance in roleplay scenarios, users should adhere to the recommended SillyTavern sampler settings and the specified Markdown character card format. The model's training on this specific structure ensures better adherence to personality and lore.