QwQ-32B-ArliAI-RpR-v2: Roleplay with Reasoning

ArliAI's QwQ-32B-ArliAI-RpR-v2 is a 32-billion parameter model designed for advanced roleplay and creative writing, leveraging a 32K token context length. It is the second iteration in the RpR (RolePlay with Reasoning) series, building upon the successful dataset and training methodologies of the RPMax series.

Key Differentiators & Capabilities

• Enhanced Reasoning for RP: Integrates reasoning processes directly into multi-turn roleplay, generated using the base QwQ Instruct model, ensuring coherent and logical progression in conversations.
• Refusal Prevention: Utilizes a "QwQ-abliterated" base to eliminate random refusals, allowing for unrestricted creative output.
• Reduced Cross-Context Repetition: Employs a unique dataset curation method to minimize repetitive phrases and tropes across different scenarios, fostering higher creativity and varied outputs.
• Optimized for Long Chats: Specifically trained to maintain consistency and quality in extended, multi-turn roleplay interactions, addressing a common limitation in other reasoning models.
• Unconventional Fine-tuning: Uses a single-epoch training approach with a higher learning rate and low gradient accumulation to prevent overfitting and encourage diverse response generation.

Ideal Use Cases

• Dynamic Roleplay Scenarios: Excels in creating engaging and non-repetitive character interactions over long conversations.
• Creative Writing & Storytelling: Suitable for generating varied narratives and avoiding common LLM "slop" or predictable writing styles.
• Applications Requiring Unrestricted Output: Beneficial for use cases where model refusals are undesirable, due to its "abliterated" base.