arogov/llama2_13b_chat_uncensored is a 13 billion parameter Llama-2 model fine-tuned by arogov using QLoRA on an uncensored Wizard-Vicuna conversation dataset. This model is designed to provide unfiltered and uncensored conversational responses, differing from standard Llama-2 chat models. It is suitable for applications requiring less restrictive or more direct AI interactions, leveraging its 4096-token context length.
Loading preview...
Overview
arogov/llama2_13b_chat_uncensored is a 13 billion parameter Llama-2 model that has been fine-tuned using QLoRA. The primary differentiator of this model is its training on the ehartford/wizard_vicuna_70k_unfiltered dataset, which is an uncensored and unfiltered conversation dataset. This training approach aims to produce a model capable of generating less restricted and more direct responses compared to conventionally moderated models.
Key Characteristics
- Base Model: Llama-2 13B, a robust foundation for conversational AI.
- Fine-tuning Method: QLoRA, enabling efficient fine-tuning on consumer-grade hardware (trained on two 24GB NVIDIA RTX 3090 GPUs).
- Training Data: Utilizes an "uncensored/unfiltered" Wizard-Vicuna conversation dataset, specifically designed to reduce inherent content restrictions.
- Training Duration: The fine-tuning process involved one epoch, taking approximately 26.5 hours.
- Prompt Style: Trained with a clear
### HUMAN:and### RESPONSE:conversational format, ensuring predictable interaction patterns.
Use Cases
This model is particularly suited for applications where the goal is to explore less constrained AI responses or for research into the effects of unfiltered training data on language models. Developers can leverage its uncensored nature for specific conversational scenarios that require directness or bypass typical content filters. The model's training code is publicly available, allowing for reproducibility and further experimentation.