Z E U S 8 B V17 abliterated V4

Introduction

ZEUS-8B-V17-ABLITERATED-V4 is a text generation model that is part of the Llama series, specifically designed to be uncensored and capable of various functionalities such as function calling and roleplay. It is developed using the Transformers library and incorporates several advanced merging techniques.

Architecture

The model is a result of merging multiple Llama models using the DARE-TIES method. The base model for this architecture is unsloth/Meta-Llama-3.1-8B-Instruct. It integrates features from several models including Orenguteng's, arcee-ai's, VAGOsolutions', and DavidAU's respective Llama-3.1 implementations. The model employs a bfloat16 data type for operations.

Training

The merge process involved detailed configuration settings including layer range specifications and parameter adjustments like density and weight filtering. A random seed of 145.0 was used for ensuring consistency in model behavior. The model's configuration allows it to perform tasks across diverse layers, leveraging the strengths of each constituent model.

Guide: Running Locally

1. Installation: Ensure you have the Hugging Face Transformers library installed. You can do this via pip:

   pip install transformers
   

2. Clone the Model: Use the Hugging Face model hub to download the ZEUS-8B-V17-ABLITERATED-V4:

   from transformers import AutoModelForCausalLM, AutoTokenizer
   
   model_name = "T145/ZEUS-8B-V17-abliterated-V4"
   model = AutoModelForCausalLM.from_pretrained(model_name)
   tokenizer = AutoTokenizer.from_pretrained(model_name)
   

3. Inference: Utilize the model for text generation tasks by encoding inputs and decoding outputs:

   inputs = tokenizer("Your input text here", return_tensors="pt")
   outputs = model.generate(**inputs)
   print(tokenizer.decode(outputs[0]))
   

4. Hardware Suggestions: For optimal performance, consider using cloud-based GPUs such as those available on AWS, Google Cloud, or Azure, as the model requires significant computational resources.

License

The model operates under the Llama3.1 license, which must be reviewed and adhered to for any use or distribution of the model.