---
language:
- en
- zh
task_categories:
- translation
size_categories:
- 10M<n<100M
license: other
license_name: upstream-source-dependent
license_link: https://huggingface.co/datasets/DALabCommunity/Haidass-translate-Datasets/blob/main/LICENSE_NOTICE.md
pretty_name: Haidass Translate Dataset
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train-*.jsonl.gz
---

# Haidass Translate Dataset

An English–Chinese translation dataset containing **18,935,335 directional examples** from 21 source pools. See [Data Sources and Composition](DATA_RECIPE.md) for the complete source breakdown.

| Direction | Examples |
|---|---:|
| English to Chinese | 10,106,394 |
| Chinese to English | 8,828,941 |
| Total | 18,935,335 |

## Usage

```python
from datasets import load_dataset

dataset = load_dataset(
    "DALabCommunity/Haidass-translate-Datasets",
    split="train",
    streaming=True,
)
example = next(iter(dataset))
```

The training split reads only `data/train-*.jsonl.gz`. Each record contains `src` (source corpus), `direction`, `instruction`, `input`, `output`, and identifier/provenance fields. Counts refer to directional examples, not bilingual pairs.

The supplementary files under `provenance/` overlap with the training split and should not be appended to it. File sizes, record counts and SHA-256 checksums are available in `PACKAGE_MANIFEST.json` and `SHA256SUMS`.

## Licensing

Source-specific terms apply. See [License Notice](LICENSE_NOTICE.md).