---
pretty_name: AstheralDataset
language: [en, de, fr, es, it, pt]
task_categories: [translation, text-generation]
license: other
license_name: source-corpus-terms
license_link: https://huggingface.co/datasets/Helsinki-NLP/opus-100
configs:
- config_name: default
  data_files:
  - {split: train, path: data/train.parquet}
  - {split: validation, path: data/validation.parquet}
  - {split: test, path: data/test.parquet}
- config_name: original
  data_files:
  - {split: train, path: original/train.parquet}
  - {split: validation, path: original/validation.parquet}
  - {split: test, path: original/test.parquet}
---
# AstheralDataset

11289 examples covering English translation into German, French, Spanish, Italian and Portuguese, plus coding, stories, sentence exercises and safety alignment.

## Splits

| Configuration/split | Rows |
|---|---:|
| default/train | 1861 |
| default/validation | 104 |
| default/test | 100 |
| original/train | 8303 |
| original/validation | 466 |
| original/test | 455 |

## Text quality

Unicode and line endings are normalized. Translation selection checks language identity, sentence structure, numeric consistency and multilingual semantic alignment. Python examples parse and compile; sentence exercises have two complete sentences. Duplicate normalized text values are removed across splits.

[Quality report](quality_report.json) · [Examples](PREVIEW.md) · [Safety examples](SAFETY_ALIGNMENT.md)

## Sources

Translation pairs: [OPUS-100](https://huggingface.co/datasets/Helsinki-NLP/opus-100), revision 805090dc28bf78897da9641cdf08b61287580df9. Source corpus terms apply.

Original instruction-response examples were generated for SEE7 using Qwen2.5-1.5B-Instruct and Qwen2.5-7B-Instruct; per-row provenance is retained.