---
license: cc-by-4.0
language:
- en
- fr
- de
- ar
- zh
task_categories:
- translation
- feature-extraction
- sentence-similarity
size_categories:
- 1K<n<100K
tags:
- parallel-sentences
- multilingual
- opus-100
pretty_name: OPUS-100 5-way (en/fr/de/ar/zh)
---

# OPUS-100 5-way parallel (en / fr / de / ar / zh)

Five-way parallel sentences derived from
[`Helsinki-NLP/opus-100`](https://huggingface.co/datasets/Helsinki-NLP/opus-100).

## How it was built

OPUS-100 is English-centric and ships ~1M sentence pairs *per* `XX-en` config.
For each of the four target languages we built a dictionary
`english_sentence -> non_english` from the `train`, `validation`, and `test`
splits of the relevant config (`en-fr`, `de-en`, `ar-en`, `en-zh`), then took
the **intersection of English keys** across all four. Each row therefore has
the same English anchor sentence aligned to a French, German, Arabic, and
Chinese translation that originally appeared in OPUS-100.

Cleaning per cell:

- strip surrounding whitespace, collapse runs of whitespace
- drop rows where any cell is shorter than 3 or longer than 1000 characters
- drop control characters
- on duplicate English keys within a single source config, keep the first occurrence
- final dedup on the `en` column

## Schema

| column | language        |
|--------|-----------------|
| `en`   | English (anchor)|
| `fr`   | French          |
| `de`   | German          |
| `ar`   | Arabic          |
| `zh`   | Chinese         |

Single split: `train` (9329 rows).

## Usage

```python
from datasets import load_dataset
ds = load_dataset("nirmalendu01/opus100-en-fr-de-ar-zh", split="train")
print(ds[0])
```

## License & citation

Inherits the OPUS-100 license (CC BY 4.0). Please cite the original OPUS-100
paper (Zhang et al., 2020, "Improving Massively Multilingual Neural Machine
Translation and Zero-Shot Translation") if you use this dataset.