EVA-UNIT-01/EVA-Qwen2.5-72B-v0.1

EVA-Qwen2.5-72B-v0.1: Roleplay and Storywriting Specialist

EVA-Qwen2.5-72B-v0.1 is a 72.7 billion parameter model, developed by Kearm, Auri, and Cahvay, specifically fine-tuned for roleplay (RP) and storywriting applications. This model is a full-parameter finetune of the Qwen2.5-72B base architecture, building upon and significantly expanding the data mixture used in Celeste 70B 0.1.

Key Capabilities & Features

• Specialized Finetuning: Optimized for creative text generation, particularly roleplay and story creation.
• Enhanced Coherence: Version 0.1 features reprocessed datasets and an readjusted training configuration, leading to significant improvements in instruction following, long context understanding, and overall narrative coherence compared to its predecessor.
• Extensive Training Data: Trained on a diverse mixture of synthetic and natural datasets, including:
  • Celeste 70B 0.1 data mixture (excluding Opus Instruct subset)
  • Kalomaze's Opus_Instruct_25k (filtered)
  • Subsets from ChatGPT-4o-WritingPrompts and Sonnet3.5-Charcards-Roleplay by Gryphe
  • Synthstruct and SynthRP datasets by Epiculous
  • Filtered subsets from Dolphin-2.9.3 (not_samantha, systemchat)
• Long Context: Supports a context length of 131072 tokens, beneficial for extended roleplay scenarios and complex story arcs.
• ChatML Format: Uses the ChatML prompt format for interaction.

Recommended Usage

This model is ideal for applications requiring highly creative, coherent, and context-aware text generation in roleplay and storywriting domains. Users are recommended to use specific sampler values for optimal performance:

• Temperature: 1
• Min-P: 0.05
• Top-A: 0.2
• Repetition Penalty: 1.03

SillyTavern presets for context and instruct/system prompts are also available for enhanced roleplay experiences.