Datasets:

YanZhanPKU
/

Entropy-Valley-Datasets

license: other
license_name: wmt-research-terms
license_link: https://www.statmt.org/wmt19/translation-task.html
task_categories:
  - translation
language:
  - en
  - zh
  - de
tags:
  - machine-translation
  - masked-diffusion
  - diffusion-language-model
  - wmt
  - length-prediction
  - arxiv:2608.22274
pretty_name: Entropy-Valley Datasets
size_categories:
  - 100K<n<1M
configs:
  - config_name: enzh
    data_files:
      - split: train
        path: enzh_train.jsonl
      - split: validation
        path: enzh_dev.jsonl
      - split: test
        path: wmt22_enzh_test.jsonl
  - config_name: ende
    data_files:
      - split: train
        path: ende_train.jsonl
      - split: validation
        path: ende_dev.jsonl
      - split: test
        path: wmt22_ende_test.jsonl
  - config_name: challenge_coverage_zh
    data_files:
      - split: test
        path: challenge_sets/challenge_coverage_zh.jsonl
  - config_name: challenge_numbers
    data_files:
      - split: test
        path: challenge_sets/challenge_numbers.jsonl
  - config_name: challenge_named_entities
    data_files:
      - split: test
        path: challenge_sets/challenge_named_entities.jsonl
  - config_name: challenge_dates
    data_files:
      - split: test
        path: challenge_sets/challenge_dates.jsonl
  - config_name: challenge_enumeration
    data_files:
      - split: test
        path: challenge_sets/challenge_enumeration.jsonl
  - config_name: challenge_fertility_mismatch
    data_files:
      - split: test
        path: challenge_sets/challenge_fertility_mismatch.jsonl

Entropy-Valley Datasets

📄 Paper (arXiv:2608.22274) | 💻 GitHub | 🤗 Models

This repository contains every data file read by Entropy-Valley (EV), the training-free target-length selector for masked diffusion machine translation introduced in "Length-Adaptive Decoding for Masked Diffusion Machine Translation" (EMNLP 2026 Main Conference).

Masked diffusion language models decode by filling a fixed-size canvas, so the target length must be chosen before denoising starts. EV probes $K{=}5$ candidate canvas lengths with one all-mask forward pass each and picks the one with the lowest mean predictive entropy. These files are the LoRA-SFT training data, the WMT22 evaluation sets, and the diagnostic challenge subsets used to measure that decision.

Dataset Summary

Config	Split	Rows	Fields	Role in the paper
`enzh`	train	200,000	`en`, `zh`	LoRA-SFT data for both En→Zh and Zh→En
`enzh`	validation	2,000	`en`, `zh`	Held-out dev split (not used to tune EV)
`enzh`	test	2,037	`en`, `zh`	Official WMT22 News test set; main results table
`ende`	train	200,000	`en`, `de`	LoRA-SFT data for the matched-scale En→De check
`ende`	validation	2,000	`en`, `de`	Held-out dev split
`ende`	test	2,037	`en`, `de`	Official WMT22 News test set (En→De)
`challenge_coverage_zh`	test	400	`en`, `zh`, `challenge_categories`, `idx`	Union subset for the reveal-order diagnostic
`challenge_numbers`	test	200	`en`, `zh`, `challenge_category`, `idx`	Sentences containing digit runs
`challenge_named_entities`	test	200	`en`, `zh`, `challenge_category`, `idx`	Capitalised multi-token spans
`challenge_dates`	test	63	`en`, `zh`, `challenge_category`, `idx`	Year / month / weekday tokens
`challenge_enumeration`	test	131	`en`, `zh`, `challenge_category`, `idx`	Comma-separated lists of ≥3 items
`challenge_fertility_mismatch`	test	14	`en`, `zh`, `challenge_category`, `idx`	Source–target length-ratio outliers

Every file is JSONL: one example per line, one JSON object per line, no nesting.

Note on direction. Zh→En reuses the same enzh files — lang_pair: zh-en in the code simply swaps which key is source and which is target. There is no separate zhen config.

Construction Protocol

Training and dev splits

Built from the HuggingFace wmt19 corpus with ladit/data/prepare_mt_data*.py:

En↔Zh — load_dataset("wmt19", "zh-en", split="train"). Filter: 5 <= len(en.split()) <= 200 and 2 <= len(zh) <= 600 (English whitespace words, Chinese characters).
En→De — load_dataset("wmt19", "de-en", split="train"), shuffled with seed=42 before iteration. Filter: 5 <= len(en.split()) <= 200 and 5 <= len(de.split()) <= 200.

After filtering, pairs are shuffled deterministically (random.seed(42) in the released scripts); the first 200,000 become the train split and the next 2,000 become the dev split.

Only single-side length filters are applied. No language-ID classifier and no source–target length-ratio filter.

What the dev split is not used for. The WMT19 development pairs are not used to set the reported EV ratio grids or any decoding hyperparameter. The grids reported in the paper are direction-specific and were informed by diagnostic range comparisons on WMT22 subsets; the WMT19 training-corpus medians (0.80 / 1.24 / 1.48) serve only as a reference scale. Once chosen, a grid is fixed and used for every sentence and every length method in that direction.

Test sets

The official WMT22 News Translation test sets, $N{=}2{,}037$ per direction.

En↔Zh — sources and references from wmt-conference/wmt22-news-systems, combined into the JSONL schema above. En→Zh and Zh→En reuse the same file with roles swapped.
En→De — fetched via sacreBLEU's get_source_file("wmt22", "en-de") and get_reference_files("wmt22", "en-de").

Challenge subsets

Derived from wmt22_enzh_test.jsonl by ladit/data/build_challenge_sets.py with --max_per_category 200. Subsets are not mutually exclusive: challenge_coverage_zh is the 400-sentence union used for the reveal-order diagnostic, and each per-category file carries its trigger in challenge_category. idx is the 0-based row index into wmt22_enzh_test.jsonl, so every challenge example can be traced back to its source row.

Note that the per-category counts here are the files as built with the 200-cap; the paper's reveal-order table reports the category composition within the 400-sentence union subset, which is smaller for the capped categories.

Usage

from datasets import load_dataset

enzh  = load_dataset("YanZhanPKU/Entropy-Valley-Datasets", "enzh")
ende  = load_dataset("YanZhanPKU/Entropy-Valley-Datasets", "ende")
nums  = load_dataset("YanZhanPKU/Entropy-Valley-Datasets", "challenge_numbers", split="test")

print(enzh)
# DatasetDict({train: 200000, validation: 2000, test: 2037})
print(enzh["test"][0])
# {'en': "I'm sorry that your order is running late.", 'zh': '很抱歉，您点的餐可能会晚到一会。'}

Or download the raw JSONL tree, which is the layout the released code expects:

huggingface-cli download YanZhanPKU/Entropy-Valley-Datasets --repo-type dataset --local-dir ./data

Then run the full pipeline from the code repository:

bash scripts/train_and_eval.sh configs/enzh.yaml 42

Rebuilding from Source

Nothing here is hand-annotated — every file is reproducible from the public WMT releases:

python -m ladit.data.prepare_mt_data      --output_dir data --num_train 200000 --seed 42
python -m ladit.data.prepare_mt_data_ende --output_dir data --num_train 200000 --seed 42
python -m ladit.data.build_challenge_sets --input data/wmt22_enzh_test.jsonl \
    --outdir data/challenge_sets --max_per_category 200

Licensing and Provenance

The paper's Resource availability paragraph (Appendix A) states that the processed experiment datasets and trained LoRA adapters are released here; this repository is that release.

These files are derived subsets of the public WMT19 training corpora and the public WMT22 News Translation test sets, redistributed here in a reformatted JSONL layout so that the paper's results are reproducible without re-running the sampling pipeline. They are not a new corpus, and no raw WMT distribution is mirrored in full.

Use is subject to the original WMT terms, which permit use for machine-translation research. Source references:

WMT19 translation task: https://www.statmt.org/wmt19/translation-task.html (HF mirror: wmt/wmt19)
WMT22 news systems and test sets: https://github.com/wmt-conference/wmt22-news-systems

If you are a rights holder for any WMT19 sub-corpus and want a portion removed, please open an issue on the code repository.

Privacy. The WMT news benchmarks are public and may contain names or events already present in public news text. No private user data was collected, and no attempt was made to identify individuals.

Citation

@inproceedings{zhan2026lengthadaptive,
  title         = {Length-Adaptive Decoding for Masked Diffusion Machine Translation},
  author        = {Zhan, Yan and Hou, Mengkai and Zhang, Wanting and Gao, Zhijun},
  booktitle     = {Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing (EMNLP)},
  year          = {2026},
  eprint        = {2608.22274},
  archivePrefix = {arXiv},
  primaryClass  = {cs.CL},
  url           = {https://arxiv.org/abs/2608.22274}
}

Please also cite the underlying WMT shared tasks:

@inproceedings{barrault-etal-2019-findings,
  title     = {Findings of the 2019 Conference on Machine Translation (WMT19)},
  author    = {Barrault, Lo{\"i}c and Bojar, Ond{\v{r}}ej and Costa-juss{\`a}, Marta R. and others},
  booktitle = {Proceedings of the Fourth Conference on Machine Translation},
  year      = {2019}
}

@inproceedings{kocmi-etal-2022-findings,
  title     = {Findings of the 2022 Conference on Machine Translation (WMT22)},
  author    = {Kocmi, Tom and Bawden, Rachel and Bojar, Ond{\v{r}}ej and others},
  booktitle = {Proceedings of the Seventh Conference on Machine Translation},
  year      = {2022}
}