N-Bot-Int/MrgrtV1-8B-merged

MrgrtV1-8B-merged: A Roleplay-Optimized 8B Model

MrgrtV1-8B-merged, developed by N-Bot-Int, is an 8 billion parameter model specifically engineered for high-quality roleplaying experiences. It is built on MuXodious/Llama-3.3-8B-Instruct-128K-PaperWitch-heresy and showcases N-Bot-Int's improved training pipeline and innovations.

Key Capabilities & Features

• Enhanced Roleplay Coherence: Trained for two weeks on Kaggle, MrgrtV1-8B demonstrates superior coherence compared to previous models, particularly in contained-roleplay scenarios.
• MEulysis-Cleaned Dataset: Utilizes a unique 100,000-entry synthetic dataset, generated from multiple capable RP models (including Mythomax and Hermes), covering diverse character types, settings, and narrative styles. This dataset includes both elaborate and taboo scenarios.
• Llama 3 Chat Formatting: The training data incorporates Llama 3 chat formatting to maximize quality and compatibility with Llama-based models.
• Uncensored Nature: The model is noted for its uncensored nature, requiring users to handle it with care and ethical considerations during fine-tuning.
• Training Methodology: Fine-tuned using Unsloth AI and Huggingface TRL on Kaggle's free tier (T4 x2 GPUs) for 3 epochs, achieving a final training loss of 1.1.

Good For

• Roleplay-Focused Applications: Ideal for scenarios requiring detailed and coherent character interactions and narrative progression.
• Contained-Roleplay: Excels particularly in structured or contained roleplay environments, though future versions aim to improve open roleplay capabilities.
• Developers Seeking Uncensored Models: Provides a base for those who require an uncensored model for specific applications, with a strong emphasis on responsible and legal use.
• Llama 3 Ecosystem Users: Optimized for use with Llama 3 formatting, ensuring better performance and integration.