EVA-Qwen2.5-14B-v0.0: Roleplay and Storywriting Specialist

EVA-Qwen2.5-14B-v0.0 is a 14.8 billion parameter model, developed by Kearm and Auri, that has undergone a full-parameter fine-tune of the Qwen2.5 architecture. It is specifically designed as a specialist for roleplay and story writing, aiming to improve versatility, creativity, and the "flavor" of generated text.

Key Capabilities & Features

• Roleplay and Storywriting Optimization: Fine-tuned on an expanded data mixture, including elements from Celeste 70B 0.1, Kalomaze's Opus_Instruct_25k, and various writing prompt and character card datasets.
• Extensive Context Window: Supports a context length of 131,072 tokens, beneficial for maintaining long-form narratives and complex roleplay scenarios.
• ChatML Prompt Format: Utilizes the ChatML prompt format for interaction.

Training Details

The model was trained using 4xA6000 GPUs for 14 hours. The training data includes:

• Celeste 70B 0.1 data mixture (excluding Opus Instruct subset)
• Kalomaze's Opus_Instruct_25k (filtered for refusals)
• Subsets of ChatGPT-4o-WritingPrompts by Gryphe and Sonnet3.5-Charcards-Roleplay by Gryphe
• A cleaned subset of shortstories_synthlabels by Auri
• Synthstruct and SynthRP datasets by Epiculous

Recommended Usage

For optimal performance, the developers recommend specific sampler values:

• Temperature: 0.7
• Top-P: 0.8
• Repetition Penalty: 1.03
  The model reportedly performs best with lower temperatures (0.8 or below) and does not favor the Min-P sampler.