Model Overview

This model, TheBloke/WizardLM-7B-V1-0-Uncensored-SuperHOT-8K-fp16, is a 7 billion parameter language model resulting from a merge of two distinct models:

• Eric Hartford's WizardLM-7B-V1.0-Uncensored: A retraining of WizardLM-7B-V1.0 with a filtered dataset to reduce refusals, avoidance, and bias, providing a more compliant, uncensored response style. It uses Vicuna-1.1 style prompts.
• Kaio Ken's SuperHOT 8K: A prototype LoRA (Low-Rank Adaptation) focused on NSFW content, featuring an extended 8K context length and no RLHF (Reinforcement Learning from Human Feedback).

Key Capabilities & Features

• Extended Context Window: Achieves an 8192-token context length, significantly enhancing its ability to handle longer inputs and maintain conversational coherence over extended interactions. This is enabled by the SuperHOT 8K merge and specific configuration settings.
• Uncensored Responses: Designed to reduce inherent model biases and refusals, offering more direct and unfiltered outputs compared to standard WizardLM versions.
• PyTorch fp16 Format: Provided in a format suitable for GPU inference and further model conversions or fine-tuning.
• Prompt Style: Utilizes Vicuna-1.1 style prompts, formatted as USER: <prompt>\nASSISTANT:.

Usage Considerations

• Context Extension: To fully utilize the 8K context, trust_remote_code=True is required during model loading, which automatically sets the scale parameter based on max_position_embeddings.
• Responsibility: As an uncensored model, users are responsible for the content generated and its implications.

Good for

• Applications requiring extended conversational memory or processing of long documents.
• Use cases where unfiltered or less biased responses are preferred.
• Developers looking for a 7B parameter model with enhanced context for experimentation or specific content generation tasks.