sss22213/Qwen3.8-27B-Heretic-JP-Roleplay-NSFW

Overview

Qwen3.8-27B-Heretic-JP-Roleplay-NSFW is a specialized 27 billion parameter model derived from Qwen/Qwen3.8-27B. It has undergone two key modifications: first, its refusal direction was removed using the Heretic tool, and second, it was fine-tuned with a QLoRA on approximately 18.8k Japanese roleplay conversations. A unique aspect of its training data is the inclusion of the model's own <think> reasoning traces, ensuring that the model maintains its internal thought process during roleplay.

Key Capabilities

• Extended Japanese Roleplay: Designed to maintain character consistency and scenario details over long, multi-turn Japanese roleplay sessions.
• Non-Refusal Content Generation: Explicitly trained to produce sexually explicit and potentially disturbing content on request without refusal, making it suitable for specific adult-oriented applications.
• Maintains Internal Reasoning: The training methodology preserved the model's <think> blocks, ensuring it continues to generate internal reasoning alongside its responses, which is crucial for complex roleplay scenarios.

Good For

• Adult-Oriented Japanese Roleplay: Ideal for applications requiring detailed, character-driven Japanese roleplay, including those involving sexually explicit or sensitive themes.
• Creative Writing & Storytelling: Can be used as a tool for generating complex narratives and dialogues in Japanese, particularly where character depth and consistent reasoning are important.

Limitations

• Adult Content: This model produces sexually explicit content and does not refuse. Users are responsible for its deployment and ensuring it is not accessible to minors or unconsenting individuals.
• Language and Task Specificity: While English and Chinese capabilities are inherited from the base Qwen model, the fine-tune was not evaluated on these languages or general assistant tasks, coding, or benchmarks. Performance drift on non-roleplay use cases should be expected.
• No Formal Evaluation: Beyond training loss and manual checks, no benchmark numbers exist for this specific checkpoint. The refusal and KL numbers mentioned in its lineage refer to the base NoRefusal model.
• Synthetic Data Influence: The model's style and potential biases are influenced by its synthetic training data, which was generated by DeepSeek-V3-0324 and the base Qwen model itself.