Datasets:
Add typed configs (v0.9) and the dataset card
Adds the v0.9 typed configs and a full dataset card, alongside the existing wmt26-grand.parquet, which is unchanged.
Added: seven typed configs, each with one split named evaluation (the name this repository's first card and the grand file's own metadata use):
| config | rows |
|---|---|
baselines |
10,753,092 |
challenge_sets |
7 |
items |
26,885 |
mt_systems |
55 |
pairs |
363,242 |
submissions |
11,297,384 |
systems |
73 |
grand stays a config of its own (it needs datasets>=4.7.0), and pairs becomes the default config.
README.md is replaced by the full card. The first card's points are kept: one row per item in grand, nested JSON payloads, a single evaluation split not partitioned into train/test sets, and raw annotation sessions kept outside the table. The card is rendered from the data, so every number in it is counted, not typed.
Checks run before this upload:
- validation: 0 failures (row counts, paper Table 2, personal data, schemas, a full round trip of every run against the grand file, item and pair texts, key integrity);
- all seven typed configs load with
datasets2.18.0, 4.0.0 and 5.1.0 with exact row counts; - builds are byte-identical.
Built from commit 2921d8c of the tooling: https://github.com/wmt-conference/wmt26-mteval/pull/1
Please review before merging: the merge is the release step.
changes to configs, card and readme. With the same parquet file structure which has been tested by Stefano on task1 data codebase - this pull request just applied configs and changes to README to explain a user how to use the data. Splits are based on configs and the original parquet file has not been modified to let the corresponding codebases work.
lgtm