Contextual Fine-Tuning of Language Models with Classifier-Driven Content Moderation for Text Generation

Area of Science:

• Artificial Intelligence
• Natural Language Processing
• Child Development

Background:

• Ensuring children's content appropriateness is vital in the digital age.
• Automated text generation (e.g., Large Language Models - LLMs) necessitates effective content filtering tools.
• Existing methods struggle with the nuances of children's literature.

Purpose of the Study:

• To develop a robust framework for generating and classifying children's stories based on suitability.
• To bridge the gap in existing content moderation tools for children's literature.
• To leverage fine-tuned LLMs for age-appropriate content creation and classification.

Main Methods:

• Fine-tuning LLMs (LLaMA, Mistral, Zephyr) for contextual story generation.
• Utilizing a BERT-based classifier for content suitability assessment.
• Evaluating generated stories using ROUGE, METEOR, and BERT Scores.

Main Results:

• Fine-tuned Mistral-7B and Zephyr-7B-Beta models showed significant improvements over base models in story generation quality.
• The BERT Classifier achieved high precision (0.95) and recall (0.97) in identifying unsuitable content.
• Fine-tuned models generated content more aligned with human standards.

Conclusions:

• Advanced LLMs offer a promising approach for generating age-appropriate children's stories.
• The developed framework enhances content moderation strategies for child-safe digital environments.
• This research has implications for educational technology, content curation, and parental control systems.