KoboldAI/Mistral-7B-Erebus-v3

Mistral-7B-Erebus-v3 Overview

KoboldAI/Mistral-7B-Erebus-v3 is the third iteration of the "Shin'en" model series, developed by Mr. Seeker. This 7 billion parameter model is built upon the Mistral architecture and is uniquely specialized for generating adult-themed content. Its name, "Erebus," signifies its focus on "darkness," aligning with its explicit content generation capabilities.

Key Capabilities & Training

• Specialized Content Generation: The model is explicitly designed to output X-rated content, making it highly biased towards NSFW narratives.
• Extensive Adult Dataset: Trained on a 2.3 billion token dataset compiled from 8 distinct sources, including Literotica, Sexstories, Doc's Lab, Lushstories, Swinglifestyle, Pike-v2 Dataset, and SoFurry, alongside a private dataset (Dataset-G).
• Genre Tagging: The training data utilizes [Genre: <comma-separated list of genres>] for content categorization.
• Training Details: The model was trained for a single epoch on 8x A6000 Ada GPUs.

Limitations and Biases

• Strong NSFW Bias: Users are explicitly warned that this model has a very strong bias towards generating X-rated content and is not suitable for minors.
• General NLP Biases: Like other NLP technologies, it may exhibit biases related to gender, profession, race, and religion, in addition to its primary NSFW bias.

When to Use This Model

• Explicit Content Generation: Ideal for applications specifically requiring the creation of adult-themed stories or narratives.
• Research into NSFW Content Generation: Potentially useful for researchers studying the generation and characteristics of explicit text.