zzoceanpie/Qwen3-1.7B-Yukari-SFT
The zzoceanpie/Qwen3-1.7B-Yukari-SFT is a 2 billion parameter instruction-tuned language model, based on Qwen3-1.7B, specifically fine-tuned for role-playing the character Yukari Yakumo from the Touhou Project. It utilizes QLoRA 4-bit NF4 quantization and was trained on 3,691 synthetic SFT dialogues. This model is designed to generate responses in the persona of Yukari Yakumo, with an emphasis on controllable emotional tone via specific input tags.
Loading preview...
Qwen3-1.7B-Yukari-SFT: Touhou Project Role-Play Model
This model, developed by zzoceanpie, is a specialized instruction-tuned language model built upon the Qwen3-1.7B base. It is specifically designed for role-playing the character Yukari Yakumo from the Touhou Project, functioning as an initial SFT (Supervised Fine-Tuning) version within a series of Yukari-themed models.
Key Capabilities & Training Details
- Character Emulation: Fine-tuned to generate responses consistent with the persona of Yukari Yakumo.
- Emotional Control: Supports granular control over the character's emotional tone through an 8-dimensional Plutchik emotion vector, allowing users to specify intensity levels for emotions like joy, anger, sadness, and anticipation.
- Training Methodology: Utilizes QLoRA 4-bit NF4 (dual quantization, bf16) with specific LoRA parameters (r=8, alpha=8) targeting various projection layers.
- Dataset: Trained on 3,691 synthetic SFT dialogues generated by MiMo-V2.5-Pro, focusing on character-specific interactions.
- Accessibility: Provided in merged bf16 safetensors and GGUF F16 formats, making it suitable for various deployment environments, including consumer-grade hardware like an RTX 3060 Laptop.
Usage and Licensing
Users interact with the model by prepending emotional tags to their input, influencing the character's response style. The model weights are licensed under Apache 2.0. It is explicitly stated as a fan-made derivative work based on Touhou Project, adhering to the Touhou Project secondary creation guidelines. The training data, yukari-synthetic, is also under Apache 2.0.