---
language:
- tok
task_categories:
- text-generation
size_categories:
- 10K<n<100K
---

# Toki Pona Pretrain

Monolingual Toki Pona texts for language model pretraining.

## Overview

| Split | Rows |
|-------|------|
| train | 15,868 |
| validation | 783 |

Each text is 2-3 sentences long. Sources: Toki Pona Wikipedia articles, [lipu kule](https://lipukule.org/), and [poki lapo](https://huggingface.co/datasets/NetherQuartz/poki-lapo).

These are texts that were rejected by LaBSE quality validation (cosine similarity < 0.70 with their machine translations), making them unsuitable for parallel training but still useful as monolingual pretraining data.

## Fields

- `text` — Toki Pona text
- `source` — origin file identifier

## Usage

```python
from datasets import load_dataset
ds = load_dataset("NetherQuartz/tokipona-monolingual")
```

## Split method

Deterministic 95/5 train/validation split by MD5 hash of the text.

## Sources & License

This dataset is derived from multiple sources with different licenses:
- **Toki Pona Wikipedia** (CC BY-SA)
- **[lipu kule](https://lipukule.org/)** / **[poki lapo](https://huggingface.co/datasets/NetherQuartz/poki-lapo)** (mixed licenses, see source dataset)
- Machine translations generated by **translategemma-12b-it** and **ilo-toki-gemmax2-28-9b**

No single license covers the entire dataset. Please check individual source licenses before redistribution.