zerofata/Q3.5-BlueStar-27B
zerofata/Q3.5-BlueStar-27B is an experimental 27 billion parameter language model based on Qwen 3.5, fine-tuned for conversational assistant tasks and roleplay. It supports both 'thinking' and 'non-thinking' modes, with reduced censorship in the thinking mode. The model was trained using Supervised Fine-Tuning (SFT) on approximately 23 million tokens, including Gemini synthetic data, with a focus on reasoning for creative assistant tasks.
Loading preview...
Model Overview
zerofata/Q3.5-BlueStar-27B is an experimental 27 billion parameter model, an SFT (Supervised Fine-Tuning) variant of Qwen 3.5. It is specifically designed and optimized for conversational assistant tasks and roleplay (RP), demonstrating creative capabilities.
Key Characteristics
- Base Model: Qwen 3.5 27B.
- Training: Fine-tuned using SFT on approximately 23 million tokens, including 12 million trainable tokens. The dataset incorporates Gemini synthetic data, replacing lower-quality datasets.
- Reasoning Capabilities: About 10% of the training dataset included reasoning for creative assistant tasks, which has generalized well across the model and reduces token usage for 'thinking' processes.
- Censorship: The 'thinking' mode has reduced censorship compared to the original Qwen instruct model.
- Performance Notes: The model is noted for its creativity, though it may occasionally exhibit repetition or 'brainfarts', which the developer suggests might be addressed with future DPO (Direct Preference Optimization) passes.
Recommended Usage
This model is particularly well-suited for:
- Conversational AI: Engaging in natural and creative dialogue as an assistant.
- Roleplay Scenarios: Excelling in interactive roleplaying applications.
- SillyTavern Integration: Specific recommended settings are provided for optimal performance in SillyTavern, including formats for actions, dialogue, and thoughts, as well as sampler configurations (Temp: 0.8, MinP: 0.05-0.075, Rep Pen: 1.00-1.1).