E V A L La M A 3.33 70 B v0.0

Introduction

EVA LLAMA 3.33 70B V0.0 is a model specialized for role-playing and story writing, derived from a full-parameter finetune of Llama-3.3-70B-Instruct. It leverages a diverse mixture of synthetic and natural data to enhance creative capabilities.

Architecture

The model is built using the Llama framework by Meta, specifically the Llama-3.3-70B-Instruct version. It incorporates advanced features such as LigerPlugin integrations and optimizations like fused linear cross-entropy, which contribute to its performance in creative text generation tasks.

Training

The model was trained using a combination of datasets, including Celeste 70B 0.1 data mixture, Kalomaze's Opus_Instruct_25k, and subsets from ChatGPT-4o-WritingPrompts and Sonnet3.5-Charcards-Roleplay. Training was conducted over 10 hours on 8 H100 SXM GPUs.

Guide: Running Locally

1. Ensure prerequisites: Install required libraries such as Transformers and Hugging Face CLI.
2. Download the model: Clone the model repository from Hugging Face.
3. Set up environment: Configure your environment to use GPUs, preferably cloud GPUs like NVIDIA's A100 or H100 for optimal performance.
4. Run the model: Use the provided configuration files to execute the model, ensuring the compatibility of your hardware and software setup.

Suggested Cloud GPUs

• NVIDIA A100
• NVIDIA H100

License

The model is governed by the Llama 3.3 Community License Agreement. It is available for personal, research, and commercial use, with a restriction against use by Infermatic Inc and its associates. More details can be found in the license file.