AKeigo/Llama3.1_StoryGeneration
AKeigo/Llama3.1_StoryGeneration is an 8 billion parameter Llama 3.1-based model fine-tuned specifically for story generation. It underwent Supervised Fine-Tuning (SFT) on writing prompts and Reinforcement Learning (RL) using GRPO with custom story quality criteria. This model excels at generating coherent, well-developed narratives with clear plot progression and polished language, making it ideal for creative writing applications.
Loading preview...
Model Overview
AKeigo/Llama3.1_StoryGeneration is an 8 billion parameter model derived from the Llama 3.1 architecture, specialized for generating high-quality stories. Its development involved two key stages: Supervised Fine-Tuning (SFT) exclusively on writing prompts, followed by Reinforcement Learning (RL) using the GRPO algorithm. The RL phase incorporated a custom set of criteria designed to evaluate and improve story quality, ranging from basic coherence to engaging plots, thematic depth, and nuanced writing techniques.
Key Capabilities
- Dedicated Story Generation: Optimized to produce complete and coherent narratives from given prompts.
- Quality-Driven Training: Utilizes a multi-tiered story quality rubric (Tier1-Tier5) during RL to enhance narrative structure, plot development, and emotional resonance.
- Structured Output: Designed to follow a specific input template for prompts and system instructions.
Recommended Usage
This model is particularly well-suited for creative writing tasks where the goal is to generate fictional stories. Users should employ the suggested system prompt: "You are a creative writing assistant. Use the provided prompt to create a coherent and well-developed story. Ensure the plot follows the prompt's guidelines, with a clear logical progression, substantial narrative development, and fluent, polished language." Recommended generation configurations include {"temperature": 0.93, "top_p": 0.85, "top_k": 0, "repetition_penalty":1.04} to achieve optimal creative outputs.