LyraNovaHeart/Midnight-Miqu-70B-v1.5-Safetensorsfix

Midnight-Miqu-70B-v1.5-Safetensorsfix Overview

This model is a 69 billion parameter DARE Linear merge, combining the strengths of sophosympatheia/Midnight-Miqu-70B-v1.0 and migtissera/Tess-70B-v1.6. It maintains the core feel and performance of Midnight Miqu v1.0 while incorporating improvements from Tess, particularly in passing specific tests without compromising writing quality. The model supports a long context of up to 32,768 tokens, similar to its Miqu predecessors.

Key Capabilities & Features

• Optimized for Creative Writing: Specifically designed for high-quality roleplaying and storytelling.
• Uncensored: Provides unrestricted content generation, with users responsible for its output.
• Long Context Support: Capable of handling up to 32K tokens with alpha_rope set to 1.
• Improved Performance: Addresses certain test failures of v1.0 while preserving writing quality.
• Sampler & Prompting Guidance: Includes detailed recommendations for sampler settings (e.g., Quadratic Sampling, Min-P) and prompting strategies (e.g., few-shot prompting, descriptive system messages) to optimize creative output.

Usage Considerations

• Personal Use Only: Due to its derivation from a leaked Mistral model, this merge is explicitly stated as suitable only for personal use, with no warranties or guarantees.
• "Warming Up" Required: May benefit from initial few-shot examples and descriptive system prompts to guide its writing style at the start of a new chat.
• Instruct Formats: Recommended for use with Vicuna or Mistral instruct formats.

Benchmarks

Evaluations on the Open LLM Leaderboard show an average score of 25.22, with specific metrics including IFEval (0-Shot) at 61.18 and BBH (3-Shot) at 38.54. More detailed results are available on the Hugging Face Open LLM Leaderboard.