Optimized Story Generation using DeepSeek LLM with Supervised Fine-Tuning

Automated story generation using Large Language Models (LLMs) has gained significant attention due to its applications in creative writing, education, and interactive systems. Despite advancements, existing LLMs often generate narratives that lack coherence, contextual alignment, and semantic accuracy, limiting their effectiveness in real-world applications. Addressing these challenges requires systematic model adaptation and optimization. This study focuses on enhancing the story generation capabilities of the DeepSeek-7B model through Supervised Fine-Tuning (SFT) and content moderation. The objective is to improve lexical overlap, contextual relevance, and narrative fluency by training the model on a curated dataset of structured narratives. The proposed methodology involves fine-tuning DeepSeek-7B using high-quality training data, optimizing for linguistic coherence and semantic alignment followed by content filtering. The model's performance is evaluated using ROUGE-1, ROUGE-2, ROUGE-L, METEOR, and BERT Score to quantify improvements over zero-shot generation. Experimental results demonstrate a 15% improvement in ROUGE-1, a 31% increase in ROUGE-2, and notable gains in METEOR and BERT scores, indicating superior content generation capabilities. These enhancements validate the effectiveness of SFT in refining narrative consistency and linguistic diversity. The findings contribute to advancing AI-driven storytelling by improving the structural and semantic integrity of generated content. This research underscores the potential of supervised adaptation in enhancing LLM performance for real-world generative applications.