Hugging Face's logo

Datasets:
IndexTeam
/
NAtIveLong

Tasks:
Translation
Modalities:
Tabular
Text
Formats:
json
Languages:
Chinese
English
Size:
< 1K
ArXiv:
Tags:
benchmark
evaluation
long-document
literary-translation
Libraries:
Datasets
pandas
Polars
License:
Dataset card Data Studio Files
xet
 Community
NAtIveLong / README.md
misono's picture
misono
Use only Index-Translate technical report as benchmark citation
f0351ff verified
preview code
|
Raw History Blame Contribute Delete
5.99 kB
metadata 
pretty_name: NAtIveLong Long-Document Translation Benchmark
license: other
license_name: component-specific-upstream-terms
license_link: https://github.com/EleanorJiang/BlonDe/blob/main/TERMS_OF_USE
language:
  - zh
  - en
task_categories:
  - translation
size_categories:
  - n<1K
tags:
  - benchmark
  - evaluation
  - long-document
  - literary-translation
configs:
  - config_name: bwb_metadata
    data_files:
      - split: test
        path: bwb/case-index.jsonl
  - config_name: guofeng_metadata
    data_files:
      - split: test
        path: guofeng/case-index.jsonl

NAtIveLong

NAtIveLong evaluates Chinese-to-English long-document translation at 4K, 8K, 16K, 32K and 64K source-token bands. It defines 84 BWB cases and 84 GuoFeng cases, using six held-out books per corpus from the upstream training corpora. These are project-defined splits, not the upstream official test splits.

What this release contains

This public release contains metadata, reconstruction information and evaluation code. It contains no BWB or GuoFeng source books or reference translations.

Corpus 4K 8K 16K 32K 64K Total Public delivery
BWB 24 24 12 12 12 84 Case index, book/chapter mapping, frozen prompt and provenance
GuoFeng 24 24 12 12 12 84 Case index, numerical extraction recipe and reconstruction script 
from datasets import load_dataset
bwb = load_dataset("IndexTeam/NAtIveLong", "bwb_metadata", split="test")
guofeng = load_dataset("IndexTeam/NAtIveLong", "guofeng_metadata", split="test")
assert len(bwb) == len(guofeng) == 84

The dataset viewer shows case metadata; it does not supply translation inputs. Do not use references, expected hashes or scoring metadata as model inputs.

Obtain official data

For BWB, obtain the official Train corpus from the BWB download and follow its terms. The released bwb/chapter-map.jsonl maps project-ordered chapters to physical upstream records, and bwb/case-index.jsonl provides frozen endpoints and hashes. This release does not include an automated BWB text reconstruction utility. See BWB details for the mapping and scoring input requirements.

For GuoFeng, obtain V1 Train through the official registration form, then run:

python guofeng/reconstruct.py --source raw/train.zh --reference raw/train.en \
  --output reconstructed/guofeng

The standard-library script verifies official input hashes and all 168 reconstructed source/reference hashes. The output directory must be new. See GuoFeng details. Reconstructed text is local material, subject to the upstream terms, and is not included in this public dataset.

Generate and evaluate

The evaluation guide documents prediction formats, installation, fixed scoring models and a two-document synthetic example. Use the frozen zh-en-document-v3 prompt and save each final translation with its case_id. Record the inference engine, model revision, reasoning setting, output limit and finish reason. The 64K band describes Chinese source length; reserve sufficient output and total-context budget for its English translation.

The primary score is SEGALE-COMET, using summary.json → five_band_macro.value: an equal-weight mean over the five length bands. Report BWB and GuoFeng separately, plus scored/total case coverage. Scores are on a 0–1 scale; chrF2 is auxiliary. Incomplete bands do not produce a valid full five-band score. Evaluation code includes upstream notices and licenses.

Limitations and provenance

Length labels use the frozen tokenizer with 5% tolerance. Windows overlap within books, and these tests have been used in model development and comparison. GuoFeng reconstruction preserves the existing undefined row in two 127-dylr cases; it does not silently repair the frozen benchmark.

release_manifest.json records the source revision and public file hashes. SHA256SUMS checks the delivered files. The original BWB text archive was intentionally excluded from this public release; BWB's metadata describes the original frozen data, not newly reselected examples.

Attribution and terms

We grant no additional rights in underlying books or translations. The evaluation code is Apache-2.0 under evaluation/LICENSE; upstream corpus terms apply separately. See LICENSE.md.

Citation

If you use this benchmark, please cite the Index-Translate technical report:

@techreport{indextranslate2026,
  author={Tianjiao Li and Mengran Yu and Chenyu Shi and Lusheng Zhang and
          Qisi Chen and Yanshan Zhou and Ji Qi and Jingying Liu and
          Yuang Feng and Ziang Cui and Tianxing Yan},
  title={Index-Translate: A Multilingual Translation Model Family --- Text, Speech, Controlled Dubbing, and Long-Document Translation},
  institution={Index LLM Team},
  year={2026},
  month={September},
  eprint={2609.40181},
  archivePrefix={arXiv},
  primaryClass={cs.CL},
  url={https://arxiv.org/abs/2609.40181}
}

Benchmark collection

Part of the Index-Translate Benchmarks collection.