ConicCat/Llama3_3-Nemo-Super-Writer-49B

ConicCat/Llama3_3-Nemo-Super-Writer-49B: Enhanced Writing and Roleplay Model

This model is a 49 billion parameter fine-tune of the Nemo Super 49B architecture, specifically optimized by ConicCat for advanced writing and roleplay applications. It addresses common limitations found in other LLMs by focusing on generating high-quality, extended text outputs.

Key Capabilities

• Improved Longform Writing: Designed to produce significantly longer text outputs, capable of generating up to 4000 words from a single prompt, ideal for narrative development or detailed content creation.
• Reduced 'AI Slop': The fine-tuning process has focused on minimizing generic or repetitive phrasing often associated with AI-generated content, leading to more natural and nuanced writing.
• Fewer 'Soft' Refusals: Users can expect a more cooperative model that is less prone to 'soft' refusals, enhancing its utility for creative and open-ended writing tasks.
• Maintained General Intellect: While specialized for writing, the model incorporates instruct data (from internlm/Condor-SFT-20K) to preserve general instruction-following abilities and intellectual capacity.

Training Data Highlights

Training involved a diverse set of datasets, including a reformatted version of the Gutenberg DPO dataset (ConicCat/Gutenberg-SFT) augmented for SFT, proprietary backtranslated books, and synthetic GLM-4.7 roleplay data. This blend of data contributes to its specialized writing prowess.

Good For

• Generating extensive creative content, such as stories, articles, or scripts.
• Engaging in detailed and nuanced roleplay scenarios.
• Applications requiring high-quality, long-form text generation with minimal AI artifacts.