The manjushv/kid-persona-young-3-4-merged is a 1.5 billion parameter Qwen2.5-based instruction-tuned language model developed by Minie AI. It is specifically fine-tuned on 50,000 real child utterances from the CHILDES corpus to accurately simulate the speech patterns of 3-4 year old children. This model excels at generating realistic child-like responses, including common grammatical errors and simplified vocabulary, making it ideal for testing children's AI products and child language research.
Loading preview...
Kid Persona Model (Age 3-4)
This model, developed by Minie AI, is a fine-tuned Qwen2.5-1.5B-Instruct LLM designed to generate realistic speech patterns of 3-4 year old children. It was trained on 50,000 real child utterances from the CHILDES corpus, the world's largest database of child language development. The fine-tuning process, using QLoRA, focused on capturing the unique linguistic characteristics of this age group, such as simplified grammar, common mispronunciations, and age-appropriate vocabulary.
Key Capabilities
- Realistic Child Speech Simulation: Generates responses that mimic how 3-4 year olds actually talk, including grammatical inversions, concrete answers, and occasional incorrect facts.
- Research-Backed Training: Utilizes data from the CHILDES corpus, ensuring authenticity in simulated child language.
- Efficient Training: Fine-tuned on an A100 in under 30 minutes for less than $2, demonstrating cost-effective development.
Good For
- Testing Children's Voice/Chat AI: Simulating realistic child inputs for product development and evaluation.
- Child Development Research: Programmatically studying language acquisition patterns.
- EdTech Development: Building educational products that can effectively handle and respond to real child speech.
- Speech Therapy Tools: Creating age-appropriate test cases and scenarios.
Important Disclaimer
This model is intended for simulation and testing purposes only and should NOT be used for direct interaction with real children or in child safety-critical applications. It generates responses as a child would, which includes incorrect grammar and answers, reflecting natural child development.