Overview

EVA-LLaMA-3.33-70B-v0.0 is a 70 billion parameter language model, a full-parameter fine-tune of Meta's Llama-3.3-70B-Instruct. Developed by Kearm, Auri, and Cahvay, this model is specifically designed as a specialist for roleplay (RP) and storywriting applications. It builds upon the Celeste 70B 0.1 data mixture, significantly expanding it to boost the model's versatility, creativity, and overall "flavor" in generative tasks.

Key Capabilities & Training

Specialized for RP/Storywriting: Optimized through a unique data mixture to excel in generating creative and engaging narrative content.
Expanded Data Mixture: Incorporates a diverse set of datasets including Celeste 70B 0.1 (minus Opus Instruct), Kalomaze's Opus_Instruct_25k (filtered), subsets of ChatGPT-4o-WritingPrompts and Sonnet3.5-Charcards-Roleplay, and Synthstruct/SynthRP datasets by Epiculous. It also includes a filtered subset from Dolphin-2.9.3.
Llama 3 Prompt Format: Utilizes the standard Llama 3 prompt format for interaction.
Training Hardware: Trained for 10 hours on 8xH100 SXM GPUs.

Recommended Usage

For optimal performance in creative generation, the developers recommend specific sampler values:

Temperature: 1
Min-P: 0.05
Repetition Penalty: 1.03

Licensing

This model is derived from Llama-3.3-70B-Instruct and is subject to the Llama 3.3 Community License Agreement and Acceptable Use Policy for Llama Materials. It is free for personal, research, and commercial use under these terms, with an additional clause prohibiting use by Infermatic Inc and its associates.