sophosympatheia/Midnight-Rose-70B-v2.0.3

Midnight-Rose-70B-v2.0.3: A Merged Model for Roleplaying and Storytelling

Midnight-Rose-70B-v2.0.3 is a 69 billion parameter language model developed by sophosympatheia, featuring a 32768 token context length. This model is the result of a sophisticated merging process involving several base models, including Llama-2-70b, WizardLM-70B-V1.0, tulu-2-dpo-70b, and dolphin-2.2-70b, utilizing DARE TIES and SLERP merge methods.

Key Capabilities and Characteristics

• Optimized for Creative Tasks: Specifically designed for roleplaying and storytelling, excelling in generating engaging and coherent narratives.
• Uncensored Output: Provides uncensored responses, offering flexibility for various creative applications.
• Complex Merging Architecture: Benefits from the strengths of its constituent models, inheriting both "spiciness" and "smarts" through its multi-stage merge process.
• Competitive General Intelligence: An IQ3_XXS quantized version of this model has shown high scores on EQBench, indicating strong general reasoning capabilities.
• Sampler and Prompting Flexibility: The model is highly responsive to various sampling parameters (e.g., Quadratic Sampling, Min-P) and prompting techniques, allowing users to fine-tune its behavior for specific needs.

Performance and Benchmarks

Evaluations on the Open LLM Leaderboard show an average score of 67.11, with notable results in:

• AI2 Reasoning Challenge (25-Shot): 70.65
• HellaSwag (10-Shot): 87.50
• MMLU (5-Shot): 69.64
• TruthfulQA (0-shot): 65.27
• Winogrande (5-shot): 81.22

Ideal Use Cases

• Roleplaying: Excellent for interactive roleplaying scenarios, generating character dialogue and actions.
• Storytelling: Suitable for creative writing, generating narrative content, and expanding story arcs.
• Content Generation: Can be adapted for various content generation tasks where creative and uncensored output is desired.