neuralnets/multilingual-tinystories

---
language:
- as
- doi
- gom
- gu
- kn
- mai
- ml
- mni
- ne
- or
- pa
- sa
- sat
- sd
- ta
- te
- ur
multilinguality:
- multilingual
size_categories:
- 10K<n<100K
task_categories:
- text-generation
- fill-mask
task_ids:
- language-modeling
pretty_name: Multilingual TinyStories
tags:
- stories
- children
- indian-languages
- indic
- low-resource
license: cc-by-4.0
---

# Multilingual TinyStories Dataset

A collection of children's stories in multiple Indian languages, generated for language model training.

## Dataset Details

### Currently Available Languages

This dataset contains stories in 17 Indic languages:

- **Assamese (`as`)**: 4,875 stories, 3,088,287 tokens
- **Dogri (`doi`)**: 4,924 stories, 2,556,071 tokens
- **GOM (`gom`)**: 4,879 stories, 2,437,488 tokens
- **Gujarati (`gu`)**: 12,856 stories, 9,858,511 tokens
- **Kannada (`kn`)**: 11,644 stories, 9,890,334 tokens
- **Maithili (`mai`)**: 4,872 stories, 2,363,974 tokens
- **Malayalam (`ml`)**: 11,216 stories, 9,742,815 tokens
- **Manipuri (`mni`)**: 4,870 stories, 71,024 tokens
- **Nepali (`ne`)**: 4,863 stories, 2,309,707 tokens
- **Odia (`or`)**: 13,004 stories, 9,506,384 tokens
- **Punjabi (`pa`)**: 13,144 stories, 9,669,977 tokens
- **Sanskrit (`sa`)**: 4,873 stories, 2,605,271 tokens
- **Santali (`sat`)**: 4,883 stories, 6,555,546 tokens
- **Sindhi (`sd`)**: 4,881 stories, 2,029,536 tokens
- **Tamil (`ta`)**: 12,860 stories, 9,840,128 tokens
- **Telugu (`te`)**: 10,924 stories, 9,865,743 tokens
- **Urdu (`ur`)**: 3,374 stories, 1,519,067 tokens

**Total stories**: 132,942
**Total tokens**: 93,909,863

> **Note**: Bengali, Marathi, and Hindi are excluded as they already have extensive resources available via the Regional TinyStories by Vizuara.

### Dataset Structure

The dataset is organized by language splits. Each split contains stories in that specific language.

```python
from datasets import load_dataset

# Load all languages
dataset = load_dataset("neuralnets/multilingual-tinystories")

# Load specific language
dataset = load_dataset("neuralnets/multilingual-tinystories", split="gu") # Gujarati

```

### Data Fields

* `text`: The story text in the respective language (native script)
* `index`: Unique identifier for each story in format `{lang_code}_{number}` (e.g., `gu_00001`, `kn_00523`)

### Usage Example

```python
from datasets import load_dataset

# Load Gujarati stories
gujarati_stories = load_dataset("neuralnets/multilingual-tinystories", split="gu")

# Print first story
print(gujarati_stories[0]["text"])
print(f"Index: {gujarati_stories[0]['index']}") # Output: gu_00000

# Load all languages
all_stories = load_dataset("neuralnets/multilingual-tinystories")
print(all_stories.keys()) # dict_keys(['gu', 'kn', 'ml', ...])

# Filter by language using index
gujarati_only = [story for story in gujarati_stories if story['index'].startswith('gu_')]

```

### Current Statistics

| Code | Language | Stories | Tokens | Status |
| --- | --- | --- | --- | --- |
| `as` | Assamese | 4,875 | 3,088,287 | ✅ Available |
| `doi` | Dogri | 4,924 | 2,556,071 | ✅ Available |
| `gom` | GOM | 4,879 | 2,437,488 | ✅ Available |
| `gu` | Gujarati | 12,856 | 9,858,511 | ✅ Available |
| `kn` | Kannada | 11,644 | 9,890,334 | ✅ Available |
| `mai` | Maithili | 4,872 | 2,363,974 | ✅ Available |
| `ml` | Malayalam | 11,216 | 9,742,815 | ✅ Available |
| `mni` | Manipuri | 4,870 | 71,024 | ✅ Available |
| `ne` | Nepali | 4,863 | 2,309,707 | ✅ Available |
| `or` | Odia | 13,004 | 9,506,384 | ✅ Available |
| `pa` | Punjabi | 13,144 | 9,669,977 | ✅ Available |
| `sa` | Sanskrit | 4,873 | 2,605,271 | ✅ Available |
| `sat` | Santali | 4,883 | 6,555,546 | ✅ Available |
| `sd` | Sindhi | 4,881 | 2,029,536 | ✅ Available |
| `ta` | Tamil | 12,860 | 9,840,128 | ✅ Available |
| `te` | Telugu | 10,924 | 9,865,743 | ✅ Available |
| `ur` | Urdu | 3,374 | 1,519,067 | ✅ Available |

## Dataset Creation

This dataset was created using language models to generate simple children's stories in various Indian languages, suitable for training small language models.

### Curation Process

1. **Generation**: Stories generated using Sarvam AI models
2. **Cleaning**: Removed emojis, English words, and formatting artifacts
3. **Native Scripts**: All stories are in their native scripts (Gujarati, Kannada, Malayalam, Devanagari, etc.)
4. **Quality**: Each story is a complete, coherent narrative suitable for children

### Index Format

Each story has a unique index in the format `{{language_code}}_{{number:05d}}`:

* `gu_00000` - First Gujarati story
* `kn_01234` - 1235th Kannada story
* `ml_00099` - 100th Malayalam story

This format allows easy identification and filtering by language.

## Use Cases

* Training small language models for Indian languages
* Multilingual language model research
* Cross-lingual transfer learning
* Educational applications
* Low-resource language modeling

## Limitations

* Stories are generated, not human-written
* May contain cultural or linguistic inaccuracies
* Not reviewed by native speakers
* Limited to simple children's story vocabulary

## Licensing

Please check individual language regulations and usage rights for your specific use case.

## Updates

Follow [@neuralnets](https://huggingface.co/neuralnets) for updates on this dataset and future projects.

## Citation

```bibtex
@dataset{multilingual_tinystories_2026,
title={Multilingual TinyStories: Indic Language Stories Dataset},
author={NeuralNets},
year={2026},
publisher={Hugging Face},
howpublished={\url{[https://huggingface.co/datasets/neuralnets/multilingual-tinystories](https://huggingface.co/datasets/neuralnets/multilingual-tinystories)}}
}

```

## Contact

For questions, issues, or contributions, please open an issue on the dataset repository.