ApocalypseParty/Qwen3.5-27B-v3-SFT-2
ApocalypseParty/Qwen3.5-27B-v3-SFT-2 is a 27 billion parameter Qwen3.5-based language model fine-tuned for roleplay and writing tasks, supporting a 32768 token context length. This model is specifically optimized for generating creative text, dialogue, and actions, with support for both 'thinking' and 'non-thinking' modes. Its training incorporated reasoning data from GLM5 and K2.5, alongside a diverse set of RP samples, making it suitable for interactive narrative applications.
Loading preview...
Overview
ApocalypseParty/Qwen3.5-27B-v3-SFT-2, also known as BlueStar v3, is a 27 billion parameter model built on the Qwen3.5 architecture. It is specifically fine-tuned for roleplay (RP) and general writing tasks, supporting a context length of 32768 tokens. This iteration, v3, incorporates additional RP reasoning data from GLM5 and K2.5, enhancing its ability to handle complex narrative interactions.
Key Capabilities
- Roleplay and Writing: Optimized for generating engaging roleplay scenarios, character dialogue, and descriptive actions.
- Flexible Reasoning: Supports both explicit 'thinking' (requiring a
<think>\nprefill) and 'non-thinking' modes for varied narrative control. - Structured Output: Recommends specific formats for actions (plaintext), dialogue ("In quotes"), and thoughts (In asterisks) for consistent output.
- Quantization Availability: Provided in GGUF format for broader compatibility and efficient deployment.
Training Details
This model was developed through Supervised Fine-Tuning (SFT) on approximately 56 million tokens. The training process involved replacing the previous Chub dataset with a version that focuses on multi-turn reasoning, ensuring more accurate and coherent long-form interactions. Additionally, a subset of 200 Gryphe RP samples was included to introduce lexical diversity. The model was trained using Axolotl, with specific LoRA configurations targeting various projection layers for efficient adaptation.