Datasets:

---

wmt

/

wmt26-mteval-data

Like 0

Follow

WMT: Workshop on Statistical Machine Translation 107

Two small changes to the dataset card. No data file changes.

1. Snapshot note. The card's baselines.variant note now also says which results in the findings paper were computed from the earlier pre_dedup_pre_enhy_fix snapshot, which only the grand config holds:

   • the "[no dedup]" rows of Table 9;
   • the organiser baselines converted for Task 3 (Tables 10-13 and 65-85).

   Anyone reproducing these should read that snapshot from grand. From the typed alt variant, six of the ten "[no dedup]" rows of Table 9 come out differently. This came out of an independent validation of main (71ccf6d), which found the data correct, record for record, and found that the organisers' code reproduces the paper from it.

2. Paper link. The citation of the pseudo-reference paper ("In the blind") now gives its arXiv URL, https://arxiv.org/abs/2609.13611, and its BibTeX entry has the arXiv eprint fields, so the Hub can list the paper on the dataset page.

Changed files: README.md and build_manifest.json (the tooling commit, now 689ce24). The 117 data files and wmt26-grand.parquet are untouched.

Checks:

• validation of the new build: 0 failures;
• data files byte-identical to main.

Built from commit 689ce24 of the tooling: https://github.com/wmt-conference/wmt26-mteval/pull/5 (the snapshot note was merged there as https://github.com/wmt-conference/wmt26-mteval/pull/4).

Please review and merge when convenient.