undefined443/wmt14-de-en-fairseq-sacrebleu

WMT14 De-En Fairseq SacreBLEU

This dataset contains a German-to-English WMT14 machine translation setup prepared for comparison with ordinary MT papers. The repository name is lowercase for Hugging Face and Unix compatibility; the formal task name is WMT14 De-En.

Dataset Structure

Parquet is the primary Hugging Face format:

data/train.parquet
data/validation.parquet
data/test.parquet

Raw parallel text is also preserved for traditional MT pipelines:

raw/train.de
raw/train.en
raw/validation.de
raw/validation.en
raw/test.de
raw/test.en

Metadata files:

manifest.json
validation_report.json

Splits

Split	Parquet	Raw source	Raw target	Pairs
train	data/train.parquet	raw/train.de	raw/train.en	4,548,885
validation	data/validation.parquet	raw/validation.de	raw/validation.en	3,000
test	data/test.parquet	raw/test.de	raw/test.en	3,003

Schema

Each parquet row has this structure:

{
  "id": "train-0000001",
  "translation": {
    "de": "Wiederaufnahme der Sitzungsperiode",
    "en": "Resumption of the session"
  }
}

The translation direction is German to English:

source: de
target: en

Source And Preprocessing

The training split comes from the Hugging Face/fairseq preprocessed WMT14 archive:

https://cdn-datasets.huggingface.co/translation/wmt_en_de.tgz

Mapping used for the train split:

wmt_en_de/train.target -> raw/train.de
wmt_en_de/train.source -> raw/train.en

Validation and test splits were generated with sacreBLEU:

sacrebleu -t wmt13 -l de-en --echo src > raw/validation.de
sacrebleu -t wmt13 -l de-en --echo ref > raw/validation.en

sacrebleu -t wmt14/full -l de-en --echo src > raw/test.de
sacrebleu -t wmt14/full -l de-en --echo ref > raw/test.en

Only line normalization was applied: bare carriage returns were replaced with spaces, and final newlines were ensured. No additional max-length filtering was applied.

Loading

After upload to Hugging Face, the parquet files can be loaded with:

from datasets import load_dataset

dataset = load_dataset("<namespace>/wmt14-de-en-fairseq-sacrebleu")
print(dataset["train"][0])

Evaluation

For WMT14 De-En test evaluation, use sacreBLEU against the canonical test set:

sacrebleu -t wmt14/full -l de-en -i hyp.en -m bleu

Here hyp.en should contain one English hypothesis per line, generated from raw/test.de or the translation.de field in data/test.parquet.

Caveats

This is not a reconstruction from the official raw WMT corpora. The train split is the Hugging Face/fairseq preprocessed version, while validation and test are materialized via sacreBLEU. Comparisons are most appropriate against papers using compatible WMT14 De-En preprocessing and sacreBLEU evaluation. Papers using different tokenization, filtering, compound splitting, BPE vocabularies, or non-wmt14/full test variants may not be exactly comparable.