EVA-UNIT-01/EVA-Qwen2.5-72B-v0.0

EVA-Qwen2.5-72B-v0.0: Roleplay and Storywriting Specialist

This model is a 72.7 billion parameter full-parameter finetune of the Qwen2.5-72B base model, developed by Kearm and Auri. It is specifically engineered to excel in roleplay and storywriting tasks, building upon the Celeste 70B 0.1 data mixture with significant expansions to improve its creative and versatile output.

Key Capabilities

• Enhanced Roleplay and Storywriting: Fine-tuned on a rich blend of synthetic and natural datasets, including Kalomaze's Opus_Instruct_25k, Gryphe's ChatGPT-4o-WritingPrompts, and Epiculous's Synthstruct and SynthRP datasets, to generate engaging and nuanced narratives.
• Creative Versatility: The expanded data mixture aims to provide a broader range of creative expression and 'flavor' in generated content.
• Large Context Window: Supports a substantial context length of 131072 tokens, beneficial for maintaining coherence in long-form roleplay and story generation.
• Optimized for Specific Use Cases: Recommended sampler values and SillyTavern presets are provided, indicating its intended use in interactive fiction and character-driven applications.

Important Considerations

• KV Cache Usage: Quantized KV cache with Qwen2.5 is not recommended and may degrade output quality; f16 is suggested for KV cache.
• Artifacting: Occasional — sequences can appear due to data normalization; this can be mitigated by banning token number 158 if penalties are too high.