Hugging Face Overview
ArliAI DS-R1-Distill-70B-ArliAI-RpR-v4-Large Overview
This model is the 70-billion parameter "Large" variant in ArliAI's RpR (RolePlay with Reasoning) v4 series, built on the deepseek-ai/DeepSeek-R1-Distill-Llama-70B base. It extends the successful RPMax dataset curation and training methodologies, focusing on high creativity and reduced repetition in roleplay and creative writing scenarios. A key innovation is the integration of reasoning capabilities into long, multi-turn chats, addressing the limitations of single-response reasoning datasets.
Key Capabilities & Differentiators
- Enhanced Roleplay and Creative Writing: Fine-tuned on a curated, deduplicated dataset to minimize cross-context repetition and impersonation, promoting a unique and varied writing style.
- Integrated Reasoning: Utilizes a novel approach to create a reasoning-enabled RP dataset, allowing the model to maintain coherence and logical progression in extended multi-turn conversations.
- Optimized for Long Chats: Increased training sequence length to 16K tokens improves awareness and memory over longer interactions.
- Unique Training Methodology: Employs a single-epoch training with a higher learning rate and low gradient accumulation, designed to prevent overfitting and encourage creative, non-repetitive outputs.
- Specific Sampler Recommendations: Best performance is achieved with simple sampler settings (e.g., Temperature 1.0, MinP 0.02, TopK 40) and high response tokens (2048+), avoiding repetition penalty samplers.
When to Use This Model
This model is particularly well-suited for applications requiring:
- Complex Roleplay Scenarios: Where consistent character portrayal and logical progression over many turns are crucial.
- Creative Story Generation: For generating varied and imaginative narratives without falling into repetitive tropes.
- Interactive Fiction and Chatbots: That need to maintain context and reasoning in extended, dynamic conversations.
Users should configure reasoning tokens (<think> and </think>) correctly for optimal performance, as detailed in the model's usage instructions.