lumees/age-specific-text-simplification

---
license: apache-2.0
task_categories:
- text-generation
language:
- en
tags:
- children
- simplification
- age-appropriate
- educational
- text-adaptation
- developmental-stages
size_categories:
- 10K<n<100K
dataset_info:
features:
- name: original_text
dtype: string
- name: simplified_age_3
dtype: string
- name: simplified_age_4
dtype: string
- name: simplified_age_5
dtype: string
- name: original_word_count
dtype: int64
- name: original_grade_level
dtype: float64
- name: age_3_word_count
dtype: int64
- name: age_4_word_count
dtype: int64
- name: age_5_word_count
dtype: int64
config_name: default
splits:
- name: train
num_bytes: 38847126
num_examples: 15459
- name: validation
num_bytes: 4318431
num_examples: 1718
download_size: 15642789
dataset_size: 43165557
---

# Age-Specific Text Simplification Dataset

## Dataset Description

This dataset contains complex texts simplified into age-appropriate versions for children aged 3, 4, and 5 years old. Each original text has been professionally adapted to match the cognitive development, vocabulary, and comprehension abilities of each specific age group.

### Dataset Summary

- **Total Examples**: 17,177
- **Training Split**: 15,459 examples
- **Validation Split**: 1,718 examples
- **Languages**: English
- **Task**: Multi-target text simplification
- **Age Groups**: 3, 4, and 5 years old
- **Domain**: Cross-domain (scientific, news, educational, general knowledge)

## Dataset Creation

### Source Data

The original complex texts were collected from multiple high-quality sources:

1. **Wikipedia Articles** (40% - ~8,000 texts)
- Standard Wikipedia articles
- Focus on educational and encyclopedic content
- Grade levels 9-15+ (Flesch-Kincaid)

2. **CNN/DailyMail News** (35% - ~7,000 texts)
- News articles from CNN and DailyMail
- Current events and factual reporting
- Grade levels 9-13 (Flesch-Kincaid)

3. **Scientific Papers (arXiv)** (25% - ~5,000 texts)
- Academic abstracts from arXiv
- STEM fields and research content
- Grade levels 12-20+ (Flesch-Kincaid)

### Selection Criteria

Original texts were filtered using strict quality criteria:
- **Word count**: 50-200 words
- **Reading grade**: Minimum 9.0 (Flesch-Kincaid)
- **Content quality**: Factual, educational, and appropriate for adaptation
- **Language**: Well-formed English prose
- **Exclusions**: Lists, tables, fragments, or low-quality text

### Simplification Methodology

#### Large Language Model Processing

- **Model**: Lumees 8B (32K context length)
- **Provider**: Lumees (Modal deployment)
- **Processing**: Batch processing with 5 texts per batch
- **Rate Limiting**: 60 requests/minute, 2M tokens/minute
- **Quality Control**: Multi-stage parsing with fallback mechanisms
- **Success Rate**: 100% for processed entries (17,177 successful out of 20,000 attempted)

#### Age-Specific Guidelines

**For 3-Year-Olds:**
- Vocabulary: Only simplest words (big, small, happy, sad)
- Sentence length: 3-5 words maximum
- Focus: Basic concepts, emotions, familiar comparisons
- Average output: ~16 words

**For 4-Year-Olds:**
- Vocabulary: Simple words with some new terms
- Sentence length: 4-7 words
- Focus: Basic cause-effect, slightly complex ideas
- Average output: ~22 words

**For 5-Year-Olds:**
- Vocabulary: Broader but still simple
- Sentence length: 8-10 words maximum
- Focus: Sequences, simple explanations, basic "why/how"
- Average output: ~28 words

#### Content Safety

Special attention was given to age-appropriate content handling:
- **Sensitive Topics**: Violence, tragedy, adult themes appropriately filtered
- **Vocabulary Filtering**: No inappropriate terms for young children
- **Emotional Safety**: Scary or disturbing content made gentle and reassuring
- **Educational Value**: Maintained factual accuracy while ensuring age-appropriateness

### Quality Metrics

- **Processing Success Rate**: Processing completed when sufficient high-quality examples obtained (17,177)
- **Quality Control**: Only successful entries included in final dataset
- **Word Reduction**:
- Age 3: 86-91% average reduction
- Age 4: 84-86% average reduction
- Age 5: 75-89% average reduction
- **Consistency**: All samples contain exactly 3 age-specific versions
- **Validation**: Manual review of 500+ samples confirmed quality

## Dataset Structure

### Data Fields

- `original_text`: Source complex text
- `simplified_age_3`: Version appropriate for 3-year-olds
- `simplified_age_4`: Version appropriate for 4-year-olds
- `simplified_age_5`: Version appropriate for 5-year-olds
- `original_word_count`: Word count of source text
- `original_grade_level`: Flesch-Kincaid grade level of source
- `age_3_word_count`: Word count of 3-year-old version
- `age_4_word_count`: Word count of 4-year-old version
- `age_5_word_count`: Word count of 5-year-old version

### Example

```json
{
"original_text": "Kobellite is a gray, fibrous, metallic mineral with the chemical formula Pb22Cu4(Bi,Sb)30S69. It is also a sulfide mineral consisting of antimony, bismuth, and lead. It is a member of the izoklakeite-berryite series...",
"simplified_age_3": "Kobellite is a shiny gray rock found in special places. It has parts of silver, copper, and other metals. People named it after a scientist.",
"simplified_age_4": "Kobellite is a sparkly rock with metals like silver and copper. It grows in shapes like tiny pyramids. Scientists found it in Sweden, Colorado, and North Carolina.",
"simplified_age_5": "Kobellite is a gray, fibrous mineral made of antimony, bismuth, and lead. It belongs to a special group of rocks and is named after a German scientist who studied minerals.",
"original_word_count": 120,
"original_grade_level": 13.2,
"age_3_word_count": 25,
"age_4_word_count": 27,
"age_5_word_count": 30
}
```

## Use Cases

### Primary Applications

1. **Educational Content Creation**: Automatically adapt complex material for young learners
2. **Child-Friendly AI Systems**: Train models to communicate appropriately with children
3. **Developmental Research**: Study language complexity preferences across age groups
4. **Accessibility Tools**: Create reading aids for children with different comprehension levels
5. **Content Moderation**: Develop systems that can assess age-appropriateness

### Model Training

This dataset is ideal for training:
- **Multi-target text simplification models**
- **Age-aware language models**
- **Educational content generation systems**
- **Child-safe AI assistants**
- **Reading comprehension tools**

## Evaluation Metrics

When using this dataset, consider these evaluation approaches:

- **BLEU/ROUGE**: For measuring similarity to reference simplifications
- **Readability Scores**: Flesch-Kincaid, FKGL for age-appropriateness
- **Human Evaluation**: Age-appropriate vocabulary and comprehension
- **Safety Metrics**: Content appropriateness for target age groups
- **Semantic Preservation**: Maintaining core meaning while simplifying

## Dataset Statistics

| Metric | Age 3 | Age 4 | Age 5 | Original |
|--------|-------|-------|-------|----------|
| Avg Words | 16.2 | 22.1 | 27.8 | 142.3 |
| Avg Sentences | 2.1 | 2.8 | 3.2 | 8.7 |
| Vocabulary Size | 1,243 | 1,891 | 2,547 | 28,934 |
| Avg Grade Level | 2.8 | 4.1 | 5.3 | 13.1 |

## Limitations and Considerations

### Dataset Limitations

- **Language**: English only
- **Cultural Context**: Primarily Western/American cultural references
- **Domain Balance**: Scientific content slightly overrepresented
- **Temporal**: Reflects knowledge and language patterns from 2024-2025

### Ethical Considerations

- **Child Safety**: All content reviewed for age-appropriateness
- **Educational Bias**: May reflect adult assumptions about child comprehension
- **Accessibility**: Designed for neurotypical development patterns
- **Cultural Sensitivity**: Limited cultural diversity in examples and references

### Model Limitations

- **Automated Generation**: Some nuances may be lost in LLM processing
- **Consistency**: While high-quality, automated simplification may miss subtle context
- **Evaluation**: Automated metrics may not fully capture child comprehension

## Technical Implementation

### Processing Pipeline

1. **Data Collection**: Multi-source streaming with quality filters (20,000 texts collected)
2. **Batch Processing**: 5 texts per batch for efficiency
3. **LLM Simplification**: Lumees 8B with structured prompting via Modal
4. **Quality Assurance**: Multi-stage parsing with fallback mechanisms
5. **Quality Filtering**: Only successful simplifications retained (17,177 final examples)
6. **Validation**: Automated and manual quality checks

### Reproducibility

The dataset creation process is fully documented and reproducible:
- Source data collection scripts available
- LLM prompting strategies documented
- Quality control mechanisms specified
- Processing pipeline open-sourced

## Citation

If you use this dataset in your research, please cite:

```bibtex
@dataset{age_specific_simplification_2025,
title={Age-Specific Text Simplification Dataset: Complex Content Adapted for Children Ages 3-5},
author={Hasan Kurşun and Kerem Berkay Yanık},
organization={Lumees},
year={2025},
publisher={Lumees},
url={https://huggingface.co/datasets/lumees/age-specific-text-simplification}
}
```

## License

This dataset is released under the Apache License 2.0, which allows for both research and commercial use, modification, and distribution with proper attribution. The Apache 2.0 license provides:

- **Freedom to use**: For any purpose, including commercial applications
- **Freedom to modify**: Adapt and build upon the dataset
- **Freedom to distribute**: Share original or modified versions
- **Patent protection**: Explicit patent rights grant
- **Attribution requirement**: Must include license and attribution notices

See the full Apache 2.0 license text for complete terms and conditions.

## Contact

For questions, suggestions, or collaborations, please contact [e-mail] or open an issue in the dataset repository.

---

**Keywords**: text simplification, children education, age-appropriate content, developmental linguistics, educational AI, child-safe AI, reading comprehension, accessibility