HY-Ug-Data

HY-Ug-Data is instruction-style translation training data for Uyghur, Chinese, and English. Each row contains a messages list with exactly one user message and one assistant message. The user asks for a translation and the assistant contains only the translated text.

Each retained bilingual pair appears exactly once in one assigned direction. A reverse copy is not generated for every pair. Direction assignment is deterministic and approximately balanced within each language pair.

Direction	Train	Validation
Chinese → Uyghur	2,766,911	109,588
Uyghur → Chinese	2,767,910	109,709
English → Uyghur	194,480	2,025
Uyghur → English	195,272	2,106
Total	5,924,573	223,428

Usage

from datasets import load_dataset

dataset = load_dataset("piyazon/HY-Ug-Data")
print(dataset["train"][0]["messages"])

The data uses this prompt, with the target language substituted:

Translate the following text into {language}. Note that you should only output the translated result without any additional explanation:
{source_text}

Apply the target model's chat template during training. Assistant-only loss is recommended when supported.

Sources and preparation

The dataset combines aligned entries from Bilqut dictionaries, aligned Chinese and Uyghur fields from Firefly-derived data, supplied translation pairs, and piyazon/uyghur_chinese_translation_clean. Dictionary metadata such as pinyin, categories, audio IDs, postcodes, formatting markers, and sign-language movement descriptions was excluded.

Preparation normalized Unicode and whitespace, merged duplicate bilingual pairs, and filtered empty, corrupted, copied, excessively repeated, badly script-mismatched, unresolved-markup, extreme length-ratio, and over-limit records. Multiple distinct translations can remain for the same source text.

The original validation sources were reserved for validation. English validation was selected deterministically from 1% of normalized English texts. Any training pair sharing normalized Chinese, English, or Uyghur text with validation was removed. The final normalized train/validation overlap is zero.

The uploaded Parquet rows were compared value-for-value against every record in the locally verified JSONL export. dataset_manifest.json contains split counts, direction counts, source JSONL hashes, Parquet file hashes, and per-shard sizes.

Limitations

Automatic structural and script checks do not establish semantic translation quality. Source spelling, translation, or alignment errors may remain. Normalized equality does not detect every paraphrase or shared substring. The maximum accepted length was 16,000 characters per side, which is not a tokenizer length.

The component sources may have different usage terms. This dataset card does not assert a unified license; users should review the terms of the applicable source data before redistribution or commercial use.