Fine-tuned LoRA adapters for the Tariikhna project: an AI-powered Islamic
educational comic platform for children (ages 6-12).
Purpose
Converts a single visual scene (from a segmented Sira passage) into a structured
comic-panel schema including a detailed image-generation prompt, following Islamic
depiction guidelines (prophets shown from behind, modest dress, child-appropriate content).
Base Model
Meta Llama 3.1 8B Instruct (4-bit quantized), fine-tuned with QLoRA via Unsloth.
Training
Method: QLoRA (rank 16, all linear projections)
Trainable params: ~42M (0.52% of 8B)
Dataset: scene -> schema pairs generated via a multi-layer pipeline from Sira text
Usage
Load with the base model and Unsloth FastLanguageModel - see project repository.
Academic Project
Graduation Project, Faculty of Computers and Artificial Intelligence, Cairo University.
