Nitral-AI/Hathor_Tahsin-L3-8B-v0.85

Hathor_Tahsin-L3-8B-v0.85 Overview

Hathor_Tahsin-L3-8B-v0.85 is an 8 billion parameter language model developed by Nitral-AI, built upon the Llama 3 8B Instruct architecture. This model is specifically designed to integrate creativity, intelligence, and robust performance, making it suitable for a variety of generative AI applications.

Key Capabilities and Training

The model underwent a rigorous training process over three epochs, utilizing a diverse and specialized dataset. This dataset includes:

• Private Roleplay Data: Enhances the model's ability to engage in dynamic and creative roleplaying scenarios.
• STEM Instructions/Dialogues: Improves its proficiency in understanding and generating responses related to science, technology, engineering, and mathematics.
• Opus Instructions: Further refines its instruction-following capabilities.
• Mixture of Light/Classical Novel Data: Contributes to its creative writing and narrative generation skills.

This training regimen aims to reduce repetitiveness, a common issue in some prior models, by building upon Hathor_Fractionate-v0.5 rather than Hathor_Aleph-v0.72.

Model Characteristics

• Parameter Count: 8 billion parameters.
• Context Length: Supports an 8192-token context window.
• Optimized for: Roleplaying, instruction-following, and creative text generation.

Available Quantizations

For broader accessibility and deployment flexibility, various quantized versions are available:

• GGUF Quantizations: Provided by Bartowski here.
• EXL2 Quantizations: Available from riveRiPH, including 5bpw, 6.3bpw, and 8bpw versions here, here, and here.

Recommended SillyTavern presets are also available here to optimize user experience.