EVA-Qwen2.5-32B-v0.0: Roleplay and Storywriting Specialist

EVA-Qwen2.5-32B-v0.0 is a full-parameter finetune of the Qwen2.5-32B model, developed by Kearm and Auri, specifically designed for roleplay and storywriting tasks. It builds upon the Celeste 70B 0.1 data mixture, significantly expanding it with additional synthetic and natural data to improve the model's versatility, creativity, and the 'flavor' of its generated text.

Key Capabilities

• Specialized for Creative Writing: Optimized for generating engaging and nuanced content for roleplay and storywriting scenarios.
• Enhanced Versatility: Trained on an expanded and diverse data mixture, including Celeste 70B 0.1, Kalomaze's Opus_Instruct_25k, and subsets of Gryphe's ChatGPT-4o-WritingPrompts and Sonnet3.5-Charcards-Roleplay, as well as Epiculous's Synthstruct and SynthRP datasets.
• Robust Base Model: Leverages the Qwen2.5-32B architecture, providing a strong foundation for complex text generation.
• ChatML Prompt Format: Utilizes the ChatML prompt format for interaction.

Good For

• Roleplaying: Excels at maintaining character consistency and generating dynamic dialogue.
• Story Generation: Ideal for crafting creative narratives, plotlines, and descriptive passages.
• Interactive Fiction: Suitable for applications requiring rich, context-aware text responses in interactive storytelling.

Note: The developers recommend against using quantized KV cache with Qwen2.5 due to potential degradation in output quality, suggesting f16 for KV cache instead.