Technical Details
How Fable Flux generates personalized children's stories using AI
Model Architecture
Fable Flux is powered by a fine-tuned version of OpenAI's GPT-OSS-20B model, specifically optimized for generating high-quality children's stories with educational value and age-appropriate content.
Model Pipeline
- Base Model: OpenAI GPT-OSS-20B
- Training Dataset: Children Stories Dataset
- Fine-tuning Method: QLoRA (Quantized Low-Rank Adaptation)
- Final Model: Fable Flux Fine-tuned Model
Training Dataset
The model was trained on a carefully curated dataset of children's stories designed to ensure:
- Age-appropriate language (3+ years)
- Positive moral lessons and educational value
- Diverse characters, settings, and themes
- Engaging storytelling with clear narrative structure
- Safe and encouraging content
Dataset Highlights:
- • 10,000+ high-quality children's stories
- • 200 diverse character archetypes
- • 100 varied settings and environments
- • Systematic diversity tracking
- • Quality validation and sentiment analysis
Serving Infrastructure
Fable Flux uses a modern, scalable infrastructure to deliver fast and reliable story generation:
Technology Stack
🚀 vLLM Engine
High-performance inference engine optimized for large language models, providing fast token generation and efficient memory usage.
☁️ Modal.com Platform
Serverless compute platform that automatically scales GPU resources based on demand, ensuring optimal performance and cost efficiency.
🎯 H100 GPUs
NVIDIA H100 Tensor Core GPUs provide the computational power for real-time story generation with minimal latency.
⚡ Next.js Frontend
Modern React framework with server-side rendering, optimized for performance and user experience.
Performance & Quality
Quality Assurance
Every generated story undergoes automated quality validation:
- ✓ Content Safety: Automated filtering for age-appropriate content
- ✓ Educational Value: Verification of positive moral lessons
- ✓ Language Complexity: Reading level analysis for target age group
- ✓ Narrative Structure: Proper story format with clear beginning, middle, and end
- ✓ Character Consistency: Logical character development and behavior
Technical Specifications
Model Details
- • Parameters: 20 billion
- • Architecture: Transformer-based
- • Precision: Mixed FP16/FP4
- • Context Length: 8,192 tokens
Infrastructure
- • GPU: NVIDIA H100 (80GB)
- • Memory: Tensor parallelism
- • Serving: vLLM 0.10.1+gptoss
- • Platform: Modal.com serverless
