kid-persona-young-3-4-merged/README.md

---
language: en
license: apache-2.0
library_name: transformers
pipeline_tag: text-generation
tags:
  - text-generation
  - qwen2
  - kid-persona
  - child-speech
  - childes
  - child-development
  - research
base_model: Qwen/Qwen2.5-1.5B-Instruct
datasets:
  - manjushv/childes-kid-persona-3-4
---

# Kid Persona Model (Age 3-4)

A fine-tuned LLM that generates realistic child speech patterns for ages 3-4, trained on real child utterances from the CHILDES research corpus.

## What This Is

This model simulates how 3-4 year old children actually talk. It was fine-tuned on 50,000 real child utterances from the [CHILDES corpus](https://childes.talkbank.org/) (Child Language Data Exchange System) — the world's largest database of child language development, spanning 40+ years of recorded parent-child conversations.

## Why It Exists

Every children's tech product tests with adults pretending to be kids. Adults type full sentences with perfect grammar. Real 3-year-olds say "me want dat cookie" and "her's gonna make." This model bridges that gap for:

- **Testing children's voice/chat AI** — simulate realistic child inputs
- **Child development research** — study language patterns programmatically
- **EdTech development** — build products that handle real child speech
- **Speech therapy tools** — generate age-appropriate test cases

## Examples

| Adult says | Model responds | Why it's realistic |
|-----------|---------------|-------------------|
| "What color is the sky?" | "it's green!" | 3-year-olds give wrong answers |
| "Can you count to five?" | "um... uh... five!" | Skips to the end with fillers |
| "Who is this man?" | "crazy" | One-word, concrete answers |
| "What time is supper?" | "at my house?" | Answers WHERE instead of WHEN |
| "Do you want juice?" | "yeah!" | Simple affirmative |
| "Can you describe the snowman?" | "I don't know I can" | Inverted grammar ("if" → missing) |

## Training Details

- **Base model**: Qwen/Qwen2.5-1.5B-Instruct
- **Method**: QLoRA (r=16, alpha=32, all-linear targets)
- **Data**: 50,000 real child utterances from CHILDES (ages 2-4)
- **Training**: 1 epoch, 28 minutes on A100
- **Cost**: Under $2
- **Loss**: 4.12 → 1.80

## Usage

```python
from transformers import pipeline

pipe = pipeline("text-generation", model="manjushv/kid-persona-young-3-4-merged")
result = pipe(
    "<|im_start|>system\nYou are a 3-year-old child. Respond naturally.<|im_end|>\n"
    "<|im_start|>user\nDo you want to play?<|im_end|>\n"
    "<|im_start|>assistant\n",
    max_new_tokens=20, temperature=0.9, do_sample=True,
)
print(result[0]["generated_text"])
```

## Live Demo

Try it: [Kid Persona Inference Space](https://huggingface.co/spaces/manjushv/kid-persona-inference)

## ⚠️ Important Disclaimer

**This model simulates child speech patterns. It is NOT a model for children to interact with.**

This model generates responses as a child would — including saying "yeah" to any question, giving wrong answers, and using incorrect grammar. This is by design: real 3-year-olds respond this way.

**This model should NOT be used to:**
- Interact with real children
- Replace child safety systems
- Generate content targeting children
- Train models that interact with children without additional safety layers

**This model IS designed for:**
- Testing and evaluating children's tech products
- Research into child language development
- Generating realistic test cases for EdTech/voice AI
- Understanding age-specific speech patterns

The training data comes from the CHILDES corpus, which contains recordings of real parent-child interactions collected under institutional review board (IRB) approval for research purposes.

## Data Source

[CHILDES](https://childes.talkbank.org/) — Child Language Data Exchange System
- License: CC-BY 4.0
- Citation: MacWhinney, B. (2000). The CHILDES Project: Tools for Analyzing Talk. 3rd Edition. Mahwah, NJ: Lawrence Erlbaum Associates.

## Built By

[Minie AI](https://minie.ai) — Building voice-first AI experiences for children.