Hermes-3-Llama-3.1-405B-Uncensored

Introduction

Hermes-3-Llama-3.1-405B-Uncensored is a fine-tuned version of the Hermes-3-Llama-3.1-405B model, designed to be uncensored. This model uses the uncensor dataset developed by Guilherme34 and is licensed under Llama 3.1.

Architecture

The model is based on the Hermes-3-Llama-3.1-405B architecture. It employs advanced techniques such as PEFT (Parameter-Efficient Fine-Tuning) and uses safe tensor formats. The model requires a specific system prompt to ensure it provides the desired uncensored outputs.

Training

Training Hardware

• Service: RunPod
• Datacenter: US-KS-2
• GPU: 4 x A100 SXM (80 GiB)
• CPU: 73 vCPU
• RAM: 1150 GiB

Training Procedure

The model was trained using the following hyperparameters:

• Learning Rate: 1e-05
• Batch Size: Train - 1, Eval - 1
• Optimizer: AdamW with cosine scheduler
• Epochs: 3
• Gradient Accumulation Steps: 4
• Distributed Type: Multi-GPU with 5 devices

The training results indicated a final train loss of approximately 0.7925 after 3 epochs.

Guide: Running Locally

To run the Hermes-3-Llama-3.1-405B-Uncensored model locally, follow these steps:

1. Install Dependencies: Ensure you have PyTorch, Transformers, Datasets, and Tokenizers libraries installed.
2. Download the Model: Use the Hugging Face transformers library to load the model.
3. Set Up System Prompt: Apply the mandatory system prompt for uncensored outputs.
4. Run Inference: Use a Python script to interact with the model.

Consider using cloud GPU services like AWS, Google Cloud, or RunPod for optimal performance, especially if your local hardware is limited.

License

The Hermes-3-Llama-3.1-405B-Uncensored model is licensed under Llama 3.1. Users are responsible for ensuring compliance with the license terms and for any content generated using the model.