taigatakano/transit-en-ja-5M

🛬 Transit-EnJa-5M 🛫

Overview

Transit-EnJa-5M is an English–Japanese translation dataset built from a subset of the C4 corpus. English source texts are filtered to be 64 tokens or fewer, then translated into Japanese using multiple large language models (LLMs).

This dataset is intended for training and evaluating machine translation systems (and related bilingual modeling tasks) under short-input constraints.

Dataset Composition

The release includes the following splits:

• Train: 5,000,000 pairs
• Validation / Eval: 400,000 pairs
• Test: 100,000 pairs

A larger version may be released in the future.

Data Creation Pipeline

1. Source selection: Samples are drawn from a subset of C4.
2. Length filtering: English inputs are filtered to ≤ 64 tokens.
3. Translation: Each English input is translated into Japanese using multiple LLMs.
4. Mechanical completeness check: We programmatically verify that all records have a translation (i.e., no missing Japanese outputs).

Data Format

Each example is a parallel pair:

• en: English text (source)
• ja: Japanese text (translation)

(Exact file format and field names may vary by hosting platform; please refer to the dataset files for the authoritative schema.)

Intended Use

• Supervised EN→JA translation training
• Short-text translation benchmarking
• Data augmentation for bilingual or multilingual models
• Evaluation of robustness under length constraints (≤ 64 tokens)

Limitations and Notes

• Translations are LLM-generated and may contain occasional errors, unnatural phrasing, or hallucinations.
• C4-derived text may include noisy or imperfect web content.
• The “completeness check” confirms translation presence, not translation quality.

Licensing

This dataset is released under ODC-By (Open Data Commons Attribution License).