wmt26-mteval-data / README.md
pranavgupta's picture dipteshkanojia's picture
Card: point to the task code; drop the errata and the comparisons with the paper (#3)
99e3315
|
Raw History Blame Contribute Delete
44.1 kB
metadata
pretty_name: WMT26 MT-Eval
task_categories:
  - translation
language:
  - arb
  - arz
  - bel
  - ces
  - deu
  - ekk
  - ell
  - eng
  - hin
  - hye
  - ind
  - isl
  - jpn
  - kaz
  - kor
  - lij
  - lld
  - rus
  - sme
  - tha
  - ukr
  - vie
  - zho
size_categories:
  - 10M<n<100M
tags:
  - machine-translation
  - mt-evaluation
  - quality-estimation
  - wmt
configs:
  - config_name: items
    data_files:
      - split: evaluation
        path: data/items/evaluation-*.parquet
  - config_name: pairs
    default: true
    data_files:
      - split: evaluation
        path: data/pairs/evaluation-*.parquet
  - config_name: submissions
    data_files:
      - split: evaluation
        path: data/submissions/evaluation-*.parquet
  - config_name: baselines
    data_files:
      - split: evaluation
        path: data/baselines/evaluation-*.parquet
  - config_name: systems
    data_files:
      - split: evaluation
        path: data/systems/evaluation-*.parquet
  - config_name: mt_systems
    data_files:
      - split: evaluation
        path: data/mt_systems/evaluation-*.parquet
  - config_name: challenge_sets
    data_files:
      - split: evaluation
        path: data/challenge_sets/evaluation-*.parquet
  - config_name: grand
    data_files:
      - split: evaluation
        path: wmt26-grand.parquet

WMT26 MT-Eval

Test data and evaluator predictions of the WMT26 Shared Task on Automated Translation Quality Evaluation Systems, the shared task on MT metrics and quality estimation (findings paper: Lavie et al., 2026; see Citation).

Participants built automatic evaluators of translation quality. The organisers gave them one test set, made of source texts and candidate translations taken from the WMT26 General Machine Translation task (General MT), and asked for contributions of four kinds. Tasks 1 to 3 asked for predictions; Task 4 asked for challenge sets:

task what it asks for columns
Task 1 error spans with a severity, and omissions t1_errors, t1_omission
Task 2 a quality score for each translation, from 0 to 100 t2_score
Task 3 whether a translation is free of errors t3_error_free
Task 4 challenge sets that probe the evaluators: five were submitted by research teams, and the two RefQuality sets are the organisers' own challenge_sets, challenge_set

The test set covers 21 official language pairs, 8 of them priority pairs that the organisers recommended to participants who could not run the whole set, and 7 challenge sets. Submissions to one task were also converted automatically to the downstream tasks in the findings paper; those conversions are not part of v0.9.

This repository holds the test set, every candidate translation, and the predictions of every participant run and every organiser baseline, as typed Parquet tables that load in one line.

  • 26,885 items in items: 8,518 official items in 21 language pairs and 18,367 challenge items.
  • 363,242 pairs in pairs: 252,923 official and 110,319 challenge.
  • Empty or whitespace-only hypotheses: 3,785 (3,689 exactly empty, 96 whitespace-only).
  • 18,021 pairs are reference candidates: the hypothesis equals the reference, up to surrounding whitespace.
  • 11,297,384 rows in submissions and 10,753,092 rows in baselines.
  • 73 evaluator runs in systems, 55 MT systems in mt_systems and 7 challenge sets in challenge_sets.

What this release contains

This is v0.9: typed data configs and documentation.

  • It contains the seven typed configs below and the organisers' original single file as grand.
  • v0.9 contains no human-annotation gold in its typed configs. The human evaluation was carried out by General MT and is public in its release. A gold config is planned for v1.0, together with the converted predictions and the results tables. The organisers' scoring code is in the tasks/ folder of the GitHub repository. The unchanged grand config still carries the organisers' raw annotation records in its ground_truth__record column; see Human annotations.

Quick start

Every config has one split, named evaluation: this is one evaluation set, not partitioned into train/test sets. The typed configs load with datasets>=2.18:

from datasets import load_dataset

pairs = load_dataset("wmt/wmt26-mteval-data", "pairs", split="evaluation")
print(pairs[0]["lp"], pairs[0]["src"], pairs[0]["hyp"], sep="\n")

pairs is the default config. submissions and baselines are large (11,297,384 and 10,753,092 rows), so you may prefer to stream them or to download one config:

subs = load_dataset("wmt/wmt26-mteval-data", "submissions", split="evaluation", streaming=True)

base = load_dataset("wmt/wmt26-mteval-data", "baselines", split="evaluation", streaming=True)
# `baselines` also holds the original, pre-deduplication snapshot (`alt`); see Warnings
base = base.filter(lambda row: row["variant"] == "official")
# scripts/download.py is in the GitHub repository, https://github.com/wmt-conference/wmt26-mteval
uv run python scripts/download.py --config submissions --out data/hf

The Parquet files can also be queried in place, for example with DuckDB:

import duckdb

base = "hf://datasets/wmt/wmt26-mteval-data/data"
duckdb.sql(f"""
    SELECT lp, system, count(*) AS n, sum(hyp_is_empty::int) AS n_empty
    FROM '{base}/pairs/*.parquet'
    WHERE collection = 'official'
    GROUP BY lp, system ORDER BY lp, system
""").show()
duckdb.sql(f"SELECT run_id, name, tasks, n_items FROM '{base}/systems/*.parquet'").show()

The grand config is the organisers' single file with JSON columns. The datasets library reads those as dicts from version 4.7.0 on (pip install "datasets>=4.7.0"); older versions cannot load this config, but they load the seven typed configs:

grand = load_dataset("wmt/wmt26-mteval-data", "grand", split="evaluation")

Configs

config rows one row per contents
items 26,885 test item Source text, reference and its type, the columns parsed from item_id, the paths of the media files and the number of systems that translated the item.
pairs 363,242 item and MT system The hypotheses: source, hypothesis, reference, whether the hypothesis is a copy of the reference (is_reference_candidate) and whether it is empty (hyp_is_empty). No gold columns in v0.9.
submissions 11,297,384 participant run, item and system What each participant's evaluator predicted: error spans, omission, score or error-free judgement, for the task it was entered in.
baselines 10,753,092 organiser baseline run, item and system The organisers' baselines in the same columns as submissions, plus variant (official or alt).
systems 73 official evaluator run The run's name, organisation, curated paper name and citation, allow-listed answers from its model card, and what it covers.
mt_systems 55 MT system Every system that has hypotheses in pairs, with its type, items and language pairs.
challenge_sets 7 challenge set Authors, citation, what the set tests, what its hypotheses and references are, and its counts.
grand 26,885 test item The organisers' lossless archive, unchanged: item fields, the raw annotation records and one JSON column per submission and baseline snapshot (needs datasets>=4.7.0).
  • Join the hypotheses to the predictions on (item_id, system), then join the predictions to systems on run_id; for baselines, filter variant == 'official' first (see Warnings).
  • baselines.variant is official for the released predictions: the deduplicated snapshot of the LLM baselines, the xCOMET baselines (whose files are identical before and after deduplication) and the segment-level metric baselines, which have a single snapshot. alt is the original, not deduplicated snapshot of the LLM baselines, with Task 1 columns only. An earlier snapshot, which the organisers label pre_dedup_pre_enhy_fix, is in the grand config only. The organisers' code reads it from grand for the "[no dedup]" rows of Table 9 and the organiser baselines converted for Task 3 (Tables 10-13 and 65-85).
  • A run may cover only some items, and then has no rows for the others; systems.n_items, lps and covers_all_priority say what it covers.
  • systems holds allow-listed model-card answers only: a fixed list of fields was copied, and nothing else was read. The links in link were volunteered by the teams for publication. name is the name the team gave its run (followed by the run id where several runs share one name), and canonical_name is the name the findings paper uses.
  • t1_instruction_fault is null in every row of this release.
  • The baseline diagnostic stage1_finish_reason, which some baseline runs record, appears only in the grand config.

How to read the data

Item identifiers

An item_id has six fields joined by the separator _###_:

<collection>_###_<source>_###_<target>_###_<domain or set>_###_<document>_###_<para>
official_###_eng_Latn_###_arz_Arab_###_news_###_brisbanetimes.com.au.377415_###_0
field column meaning
1 collection official for the main test set, challenge for a challenge set.
2, 3 src_lang_orig, tgt_lang_orig The language codes as written in the id; src_lang and tgt_lang hold the normalised codes.
4 domain or challenge_set The domain of an official item (news, social, speech, software, edu or factchecking), or the name of the challenge set (GAMBITP, FECE, QEbreak, AGGREE or LIGATUR). The two RefQuality sets keep the domain here and carry their name in the document field.
5 doc_id The source document. The document of a RefQuality item starts with the name of its set (RefQuality-PD_ or RefQuality-PE_); one that starts with RefQuality-PD_RefQuality-PE_ marks the pseudo-reference variant of an item whose alternative reference is a post-edit.
6 para_idx The zero-based paragraph index (see below).

para_idx is null for FECE and QEbreak, whose ids end in a segment hash instead of a paragraph number.

Split on the whole separator, not on #: item_id.split("_###_"). Document ids can themselves contain _, | and ..

Language codes

Codes are normalised to an ISO 639-3 language code and an ISO 15924 script, plus a region where one is needed (eng_Latn, zho_Hans, zho_Hant_TW); the original codes stay in src_lang_orig and tgt_lang_orig. The lp column is the paper's short form: the first three letters of each side, except that a Simplified Chinese target is zho_CN and Traditional Chinese (Taiwan) is zho_TW (eng-zho_CN, eng-zho_TW, zho-jpn).

Reference types

ref_type says how the reference was produced:

ref_type source data meaning
native human A human translation from scratch.
postedit postedit A post-edited machine translation.
pseudo pseudo A machine-generated pseudo-reference.
unknown undefined A reference text is present, but the source data does not record its type.
none absent, or null text and type The item has no reference text, and ref_text is null.

items.ref_type_orig keeps the type exactly as the source data spells it (human, postedit, pseudo, undefined, or null where none is recorded).

System types

system_type classifies the system behind a hypothesis:

system_type the system is
mt A machine translation system.
human_scratch Human (from scratch).
human_postedit Human (postediting).
pseudo_ref PseudoRef, the pseudo-reference.
challenge_hyp A challenge-set hypothesis such as FECE_hyp1.

Missing values

Every column may be null; the description of each column says when. A run that does not cover an item has no rows for it (there are no null rows), and a run that makes no prediction for a task has null in that task's columns. Text values are copied exactly from the source data: nothing is trimmed, re-cased or clipped.

Warnings

  • baselines holds two snapshots of some runs. 12 runs have an alt snapshot next to their official one, and its 3,469,756 alt rows share their run_id with the official rows. Counting or averaging baselines by run_id, or joining it to systems on run_id, therefore counts those runs twice. Filter variant == 'official' unless you want the pre-deduplication snapshot.

  • References are among the candidate translations of official items. In the official set, every available reference is also a hypothesis (system_type is human_scratch, human_postedit or pseudo_ref). Challenge hypotheses can equal the reference too: challenge_hyp rows can be is_reference_candidate = true. In all, 18,021 pairs have a hypothesis that equals the item's reference, up to surrounding whitespace, and carry is_reference_candidate = true. A reference-based evaluator is given the reference as well, so it sees a copy of it on those rows.

    is_reference_candidate also flags MT outputs that are identical to the reference, so it does not mark the reference systems only. By system_type, the flagged pairs are:

    • mt: 8,673 (MT outputs that equal the reference)
    • human_scratch: 2,888 (Human (from scratch), a reference system)
    • human_postedit: 1,714 (Human (postediting), a reference system)
    • pseudo_ref: 3,627 (PseudoRef, a reference system)
    • challenge_hyp: 1,119 (challenge hypotheses that equal the reference)

    To drop the reference itself, filter on system_type: remove the flagged rows of human_scratch, human_postedit and pseudo_ref. Do not drop every is_reference_candidate row: that removes real MT outputs and challenge hypotheses as well.

  • The GPT-5.6 Sol baseline at xHigh reasoning is incomplete. The findings paper says this run completed only 6 of the 8 priority pairs. In the data, its t2_score is the same placeholder value, 50.0, in every row, and its Task 1 predictions are empty for almost all of eng-sme (no row has an error span) and for most of eng-ind. Do not use its t2_score, and treat its eng-sme and eng-ind predictions as incomplete.

  • Some hypotheses are empty. There are 3,785 empty or whitespace-only hypotheses (3,689 exactly empty and 96 whitespace-only); hyp_is_empty is true for all of them. They are kept, and so are the predictions made for them.

  • RefQuality items repeat sources of official items. The two RefQuality challenge sets (RefQuality-PD and RefQuality-PE) re-use the sources of official items with a different reference, a pseudo-reference or a post-edit, to show what the reference type changes. Their ids differ from the official ones, so filter on collection before you count sources. A doc_id that starts with RefQuality-PD_RefQuality-PE_ marks the pseudo-reference variant of an item whose alternative reference is a post-edit.

  • Three language pairs have little human annotation. The share of pairs that General MT annotated is only 10.6 % for ces-vie, 17.1 % for eng-zho_CN and 19.0 % for eng-sme (findings paper, Table 2), so meta-evaluation on them rests on few judgements.

  • Items are paragraph-level. The last field of item_id is a paragraph index, and the "segments" of the findings paper are these items. A hypothesis can span several sentences, and the start and end of a Task 1 span are character offsets into that whole hypothesis.

  • Cohere CAT+ is two things. It is an MT system that has hypotheses in pairs and an evaluator that has a row in systems. Do not join pairs.system to systems.name: in the predictions, run_id names the evaluator and system the translation it judged.

  • Task 2 scores can fall outside 0-100. Some runs predict below 0 or above 100, and some wrote integers; every score is stored unchanged as float64.

  • The challenge sets differ in their references. GAMBITP has no reference. AGGREE, FECE, LIGATUR and QEbreak have a reference translation, but the source data does not record its type (ref_type is unknown) and the findings paper does not say how it was produced. The RefQuality sets have pseudo-references or post-edits.

  • Challenge sets are not scored like the official set. Their authors analysed the results with their own code. Which of the two hypotheses of a challenge item is which is not stated in the findings paper for several sets; see challenge_sets.hyp_semantics.

  • t1_instruction_fault is null in every row, and para_idx is null for FECE and QEbreak items.

Coverage of the official test set

One line per official language pair: its source items, its pairs (candidate translations), the type of its reference and whether it is a priority pair.

language pair source items pairs reference priority
ces-deu 438 13,578 native yes
ces-ukr 415 12,035 none no
ces-vie 418 12,958 native yes
eng-arz 433 12,990 postedit yes
eng-bel 354 10,974 pseudo no
eng-ces 448 13,440 postedit no
eng-deu 459 13,770 postedit no
eng-ekk 356 10,680 native yes
eng-hye 364 10,192 pseudo no
eng-ind 331 10,261 pseudo yes
eng-isl 360 11,160 native yes
eng-kaz 438 12,702 pseudo no
eng-lij 411 11,919 native no
eng-lld 376 10,152 native no
eng-rus 426 13,206 pseudo no
eng-sme 493 13,311 native yes
eng-tha 451 13,981 pseudo no
eng-ukr 406 12,180 pseudo no
eng-zho_CN 481 13,949 pseudo no
eng-zho_TW 325 9,100 pseudo no
zho-jpn 335 10,385 postedit yes
all official 8,518 252,923

The counts are those of the items and pairs configs; the release checks them against Table 2 of the findings paper. native is the "Human" reference of that table. The challenge sets are not in this table: they hold 18,367 items and 110,319 pairs.

Field glossary

The same tables are in docs/FORMAT.md, and the descriptions are stored in the schema of every Parquet file.

items

One row per test item.

column type description
item_id string Unique identifier of the test item: six fields (collection, source language, target language, domain or challenge-set name, document, paragraph or segment) joined by _###_.
collection string Either official for the main test set or challenge for a challenge set.
challenge_set string Name of the challenge set the item belongs to (GAMBITP, FECE, QEbreak, AGGREE, LIGATUR, RefQuality-PD or RefQuality-PE), or null for official items.
domain string Text domain (news, social, speech, software, edu or factchecking) for official and RefQuality items, or null for every other challenge item.
src_lang string Source language, normalised to a three-letter language code, a script and, where needed, a region (for example eng_Latn or zho_Hant_TW).
tgt_lang string Target language, normalised to a three-letter language code, a script and, where needed, a region (for example deu_Latn or zho_Hans).
src_lang_orig string Source language code exactly as it is written in item_id (for example en or eng_Latn).
tgt_lang_orig string Target language code exactly as it is written in item_id (for example de_DE or deu_Latn).
lp string Language pair as a short src-tgt code, such as ces-deu or eng-zho_CN.
is_priority bool True for official items whose language pair is one of the eight priority directions, and false for every other item, challenge items included.
doc_id string Identifier of the source document the item belongs to, the fifth field of item_id.
para_idx int32 Zero-based index of the paragraph within its document, taken from the last field of item_id, and null for FECE and QEbreak, whose item ids end in a segment hash.
src string Source text of the item, exactly as in the test set.
ref_text string Reference translation, exactly as in the test set, or null when the item has no reference.
ref_type string How the reference was produced: native (human translation from scratch), postedit (post-edited machine translation), pseudo (machine-generated pseudo-reference), unknown (a reference text is present but its type is not recorded) or none (no reference text).
ref_type_orig string Reference type exactly as recorded in the source data (human, postedit, pseudo or undefined), or null when the data records no type.
asr string Automatic speech recognition transcript of the source audio, present for speech items only.
video_path string Relative path of the source video file as recorded in the test set, present for speech items only.
screenshot_path string Relative path of the screenshot of the original post, present for some social media items only.
n_systems int32 Number of systems (machine translation, human translation, pseudo-reference or challenge hypothesis) that have a hypothesis for the item.

pairs

One row per item and MT system.

column type description
item_id string Unique identifier of the test item: six fields (collection, source language, target language, domain or challenge-set name, document, paragraph or segment) joined by _###_.
collection string Either official for the main test set or challenge for a challenge set.
challenge_set string Name of the challenge set the item belongs to (GAMBITP, FECE, QEbreak, AGGREE, LIGATUR, RefQuality-PD or RefQuality-PE), or null for official items.
domain string Text domain (news, social, speech, software, edu or factchecking) for official and RefQuality items, or null for every other challenge item.
src_lang string Source language, normalised to a three-letter language code, a script and, where needed, a region (for example eng_Latn or zho_Hant_TW).
tgt_lang string Target language, normalised to a three-letter language code, a script and, where needed, a region (for example deu_Latn or zho_Hans).
src_lang_orig string Source language code exactly as it is written in item_id (for example en or eng_Latn).
tgt_lang_orig string Target language code exactly as it is written in item_id (for example de_DE or deu_Latn).
lp string Language pair as a short src-tgt code, such as ces-deu or eng-zho_CN.
is_priority bool True for official items whose language pair is one of the eight priority directions, and false for every other item, challenge items included.
doc_id string Identifier of the source document the item belongs to, the fifth field of item_id.
para_idx int32 Zero-based index of the paragraph within its document, taken from the last field of item_id, and null for FECE and QEbreak, whose item ids end in a segment hash.
system string Name of the system that produced the hypothesis, exactly as in the source data.
system_type string Kind of system: mt, human_scratch, human_postedit, pseudo_ref or challenge_hyp.
src string Source text of the item, exactly as in the test set.
hyp string Hypothesis translation, exactly as in the source data, including empty and whitespace-only outputs.
ref_text string Reference translation, exactly as in the test set, or null when the item has no reference.
ref_type string How the reference was produced: native (human translation from scratch), postedit (post-edited machine translation), pseudo (machine-generated pseudo-reference), unknown (a reference text is present but its type is not recorded) or none (no reference text).
is_reference_candidate bool True when the hypothesis equals the reference apart from surrounding whitespace, so that it may be a copy of the reference; it marks the reference itself (a human_scratch, human_postedit or pseudo_ref row) and also any MT output or challenge hypothesis that equals it, so filter on system_type to tell them apart.
hyp_is_empty bool True when the hypothesis is empty or contains only whitespace.

submissions

One row per participant run, item and system.

column type description
run_id string Identifier of the evaluator run: the numeric submission id, or the name of the organisers' baseline (for example gemma_ref or 872013).
task string Task the run was entered for: task1 (error spans, with Task 2 scores as well for the organisers' Task 1 baselines), task2 (segment scores) or task3 (error-free judgements).
item_id string Unique identifier of the test item: six fields (collection, source language, target language, domain or challenge-set name, document, paragraph or segment) joined by _###_.
lp string Language pair as a short src-tgt code, such as ces-deu or eng-zho_CN.
system string Name of the system that produced the hypothesis, exactly as in the source data.
t1_errors list<struct<start: int32, end: int32, severity: string, category: string>> Task 1 error spans the run predicts for the translation, each with start and end offsets into the hypothesis, a severity and a category (which may be null), or null where the run makes no Task 1 prediction for the row.
t1_omission string Omission severity (minor or major) that the run predicts for the translation, or null when it predicts no omission or makes no Task 1 prediction.
t1_instruction_fault string Instruction fault that the run reports for the translation, or null when it reports none or makes no Task 1 prediction (null in every row of this release).
t2_score float64 Task 2 quality score the run predicts for the translation, copied unchanged from the source (some runs score below 0 or above 100), or null where the run makes no Task 2 prediction.
t3_error_free bool Task 3 judgement that the translation is free of errors, or null where the run makes no Task 3 prediction.

baselines

One row per organiser baseline run, item and system.

column type description
run_id string Identifier of the evaluator run: the numeric submission id, or the name of the organisers' baseline (for example gemma_ref or 872013).
task string Task the run was entered for: task1 (error spans, with Task 2 scores as well for the organisers' Task 1 baselines), task2 (segment scores) or task3 (error-free judgements).
item_id string Unique identifier of the test item: six fields (collection, source language, target language, domain or challenge-set name, document, paragraph or segment) joined by _###_.
lp string Language pair as a short src-tgt code, such as ces-deu or eng-zho_CN.
system string Name of the system that produced the hypothesis, exactly as in the source data.
t1_errors list<struct<start: int32, end: int32, severity: string, category: string>> Task 1 error spans the run predicts for the translation, each with start and end offsets into the hypothesis, a severity and a category (which may be null), or null where the run makes no Task 1 prediction for the row.
t1_omission string Omission severity (minor or major) that the run predicts for the translation, or null when it predicts no omission or makes no Task 1 prediction.
t1_instruction_fault string Instruction fault that the run reports for the translation, or null when it reports none or makes no Task 1 prediction (null in every row of this release).
t2_score float64 Task 2 quality score the run predicts for the translation, copied unchanged from the source (some runs score below 0 or above 100), or null where the run makes no Task 2 prediction.
t3_error_free bool Task 3 judgement that the translation is free of errors, or null where the run makes no Task 3 prediction.
variant string Snapshot of the baseline: official for the released predictions, or alt for the original predictions from before deduplication, with Task 1 fields only; the label is provisional and will be checked against the paper in v1.0.

systems

One row per official evaluator run.

column type description
run_id string Identifier of the evaluator run: the numeric submission id, or the name of the organisers' baseline (for example gemma_ref or 872013).
name string Name the team gave the submission in its model card, followed by the run id in brackets where several runs share that name, so that every name is unique.
canonical_name string Name under which the findings paper lists the system, curated by the organisers and never suffixed with the run id.
org string Organisation or organisations that entered the run, as written in the model card.
description string Short free-text description of the system, copied from the model card.
reference_usage string Free-text account of which kinds of reference translation the system uses and how, copied from the model card.
model_size string Size of the model behind the system, as the team wrote it in the model card.
closed_model_name string Name of the closed-source model the system calls, as written in the model card.
api_provider string Provider or host of the API through which the closed-source model is called, as written in the model card.
link string Link to the submitted system that the team volunteered for publication in its model card, or an empty string where it gave none.
underlying_model string Model or models behind the system as the findings paper reports them, curated by the organisers (— for the metrics that use no model, and not stated in the paper where the paper gives none).
paper_citation string Citation of the paper that describes the system, in the author-year form of the findings paper, or [Citation missing] where the findings paper has none.
is_baseline bool True for the organisers' baselines and false for participant submissions.
uses_reference bool True when the model card says that the system uses the provided reference translations, false when it says that it does not, and null when the answer is missing or is neither yes nor no.
ensemble bool True when the model card says that the system is an ensemble, false when it says that it is not, and null when the answer is missing or is neither yes nor no.
fine_tuned bool True when the model card says that the team fine-tuned its model, false when it says that it did not, and null when the answer is missing or is neither yes nor no.
open_weight bool True when the model card says that the system is based on open-weight models, false when it says that it is not, and null when the answer is missing or is neither yes nor no.
open_source bool True when the model card says that the submitted system is open-source, false when it says that it is not, and null when the answer is missing or is neither yes nor no.
threshold_inclusive bool True when a Task 2 score equal to t3_threshold counts as error-free, false when only a higher score does, and null when the model card says neither or both.
allow_t1_to_t2 bool True when the team allows the organisers to convert its Task 1 output into a Task 2 prediction automatically, false when it does not, and null when the model card gives no readable answer.
allow_t1_to_t3 bool True when the team allows the organisers to convert its Task 1 output into a Task 3 prediction automatically, false when it does not, and null when the model card gives no readable answer.
allow_t2_to_t3 bool True when the team allows the organisers to convert its Task 2 output into a Task 3 prediction automatically, false when it does not, and null when the model card gives no readable answer.
covers_all_priority bool True when the run has at least one row for every one of the eight priority language pairs among the official items.
t3_threshold float64 Task 2 score that separates the team's Task 2 predictions that are converted to error-free judgements from those that are not: a higher score counts as error-free, and a score equal to the threshold does so when threshold_inclusive is true. Null where the model card gives no numeric threshold.
tasks list<string> Tasks the run has predictions for: the one task of a submission, task1 and task2 for the organisers' Task 1 baselines, and task2 for the seven segment-level metric baselines.
lps list<string> Language pairs, as short codes sorted alphabetically, in which the run has at least one row, challenge sets included.
n_items int64 Number of distinct test items for which the run has at least one row.
n_pairs int64 Number of rows (one per item and system) that the run has in its prediction table.

mt_systems

One row per MT system.

column type description
system string Name of the system that produced the hypothesis, exactly as in the source data.
system_type string Kind of system: mt, human_scratch, human_postedit, pseudo_ref or challenge_hyp.
n_items int64 Number of distinct test items for which the system has a hypothesis.
lps list<string> Language pairs, as short codes sorted alphabetically, in which the system has hypotheses.

challenge_sets

One row per challenge set.

column type description
name string Name of the challenge set, as in the challenge_set column of items and pairs.
full_name string Title of the paper that introduced the challenge set, copied from the reference list of the findings paper, or a descriptive name for the organisers' reference-swap sets, which have no published title.
authors string Authors of the challenge set, as in the reference list of the findings paper.
citation string Citation of the paper that describes the challenge set, in the author-year form of the findings paper.
what_it_tests string Short description of the phenomenon that the challenge set probes, from the findings paper.
hyp_semantics string What the hypotheses of the set are, and which system names they appear under, or not stated in the paper where the paper does not say.
ref_semantics string What the reference of the set is, or not stated in the paper where the paper does not say.
lps list<string> Language pairs, as short codes sorted alphabetically, in which the set has items.
n_items int64 Number of distinct items in the challenge set.
n_pairs int64 Number of rows (one per item and system) that the challenge set has in pairs.

Personal data

  • The model cards of the participants ask for contact details and for consent. systems is built from a fixed allow-list of answers. The fields that hold the names and e-mail addresses of the submitters and the answers about data collection are not on the list, and their answers were never read. The allow-listed answers were screened too: an e-mail address inside an allow-listed answer is replaced by [email removed], and one allow-listed answer that named individuals was replaced by the organisations they belong to. The only e-mail address in the metadata is the organisers' mailing list, below. A release check looks for any other address in systems, challenge_sets, this card and the build manifest.
  • The links in systems.link were volunteered by the teams, in the model card, for publication.
  • Source texts and translations are published as they are in the WMT26 General MT test set, and are not changed or scanned. They may contain personal details that occur in the original public texts, such as social-media posts.
  • Annotator ids in the raw annotation records of grand are pseudonyms.

If you find personal data that should not be here, write to wmt-qe-metrics-organizers@googlegroups.com.

Human annotations

The human evaluation (contrastive error span annotation, cESA) was carried out by the WMT26 General MT task (Kocmi et al., 2026a; Zouhar et al., 2026a). Its data, including the raw annotation logs, are public in General MT's release:

https://github.com/wmt-conference/wmt26-general-mt/releases/tag/humeval

v0.9 contains no human-annotation gold in its typed configs; a gold config is planned for v1.0. The unchanged grand config still holds the organisers' raw annotation records in the column ground_truth__record.

The speech items point to video files and some social-media items to screenshots, in items.video_path and items.screenshot_path. These paths are relative to General MT's multimodal assets, the archive multimodal_inputs.zip in the release linked above.

Provenance

The typed configs are derived only from wmt26-grand.parquet (sha256 a61675552cebb68fb8303cf08a23fbec31e5bb4e0bd26a1fdee1a9bfafc2fb7a), the scrubbed model cards, the list of priority pairs, and the curated tables of names, language codes and challenge sets that follow the findings paper. build_manifest.json lists the sha256 and the rows of every Parquet file, and the GitHub repository lists the source files in data/MANIFEST.tsv. The code that builds and checks this repository is at https://github.com/wmt-conference/wmt26-mteval; the release process is in docs/RELEASE.md.

Citation

If you use this data, cite the findings paper:

@inproceedings{lavie2026findings,
  title = {Findings of the WMT26 Shared Task on Automated Translation Quality Evaluation Systems: Compact Open Models are Competitive Quality Evaluators},
  author = {Lavie, Alon and
            Hanneman, Greg and
            Perrella, Stefano and
            Ding, Shuoyang and
            Avramidis, Eleftherios and
            Proietti, Lorenzo and
            Lo, Chi-kiu and
            Shurtz, Ammon and
            Zerva, Chrysoula and
            Sindhujan, Archchana and
            Zouhar, Vilém and
            Kanojia, Diptesh and
            Blain, Frédéric and
            Thompson, Brian and
            Filandrianos, Giorgos and
            Menis Mastromichalakis, Orfeas and
            Kocmi, Tom and
            Gupta, Pranav},
  booktitle = {Proceedings of the Eleventh Conference on Machine Translation (WMT)},
  year = {2026},
  pages = {931--1030},
  url = {https://www2.statmt.org/wmt26/pdf/2026.wmt-1.49.pdf}
}

The data build on the following work, as the findings paper cites it.

The WMT26 General MT task, which provided the source texts, the translations and the human evaluation:

@inproceedings{kocmi2026general,
  title = {Findings of the WMT26 general machine translation shared task: Contrastive dynamic human evaluation at scale},
  author = {Kocmi, Tom and Artemova, Ekaterina and Avramidis, Eleftherios and Bawden, Rachel and Bojar, Ondřej and Dukanov, Sergey and Dvorkovich, Anton and Fishel, Mark and Freitag, Markus and Frontull, Samuel and Gowda, Thamme and Grundkiewicz, Roman and Haddow, Barry and Kharevich, Stan and Koehn, Philipp and Li, Zheng and Maillard, Jean and Monz, Christof and Murauski, Alexander and others},
  booktitle = {To appear in Proceedings of the Eleventh Conference on Machine Translation},
  year = {2026},
  publisher = {Association for Computational Linguistics}
}

The contrastive error span annotation protocol (cESA) of the human evaluation:

@misc{zouhar2026cesa,
  title = {Contrastive esa: Human evaluation of multiple translations at once},
  author = {Zouhar, Vilém and Grundkiewicz, Roman and Rajaee, Sara and Riley, Parker and Popel, Martin and Bawden, Rachel and Koehn, Philipp and Carpuat, Marine and Kocmi, Tom},
  year = {2026},
  eprint = {2607.26640},
  archiveprefix = {arXiv}
}

The pipeline that built the pseudo-references (https://arxiv.org/abs/2609.13611):

@inproceedings{kanojia2026blind,
  title = {In the blind: Building pseudo-references for MT evaluation},
  author = {Kanojia, Diptesh and Lo, Chi-kiu and Sindhujan, Archchana and Larkin, Samuel and Hanneman, Greg and Lavie, Alon},
  booktitle = {Proceedings of the Eleventh Conference on Machine Translation},
  year = {2026},
  publisher = {Association for Computational Linguistics},
  eprint = {2609.13611},
  archiveprefix = {arXiv}
}

The challenge sets (RefQuality-PD and RefQuality-PE are the organisers' own, in Section 5.2.1 of the findings paper):

@inproceedings{griesbeck2026aggree,
  title = {AGGREE: An aggregation-grounded challenge set for segment-level MT metric evaluation},
  author = {Griesbeck, Lena and Geierhos, Michaela},
  booktitle = {Proceedings of the Eleventh Conference on Machine Translation},
  year = {2026},
  address = {Budapest, Hungary},
  publisher = {Association for Computational Linguistics}
}

@inproceedings{ding2026fece,
  title = {FECE: A challenge set for evaluating fine-grained translation error correction},
  author = {Ding, Jie and Duan, Xiangyu},
  booktitle = {Proceedings of the Eleventh Conference on Machine Translation},
  year = {2026},
  address = {Budapest, Hungary},
  publisher = {Association for Computational Linguistics}
}

@inproceedings{menismastromichalakis2026gambit,
  title = {Benchmarking gender bias in machine translation evaluation metrics across occupations},
  author = {Menis Mastromichalakis, Orfeas and Filandrianos, Giorgos and Mohammed, Wafaa and Attanasio, Giuseppe and Zerva, Chrysoula},
  booktitle = {Proceedings of the Eleventh Conference on Machine Translation},
  year = {2026},
  address = {Budapest, Hungary},
  publisher = {Association for Computational Linguistics}
}

@inproceedings{mishra2026ligatur,
  title = {Translation metrics cannot judge what their tokeniser deletes: LIGATUR and AEGIS submissions to the WMT26 shared task on automated translation quality evaluation},
  author = {Mishra, Animesh and Sharma, Krishang and Khetarpaul, Sonia},
  booktitle = {Proceedings of the Eleventh Conference on Machine Translation},
  year = {2026},
  address = {Budapest, Hungary},
  publisher = {Association for Computational Linguistics}
}

@inproceedings{silchenko2026qebreak,
  title = {QEbreak at WMT26: A preregistered audit of omission and addition asymmetry in machine translation quality metrics},
  author = {Silchenko, Maksim},
  booktitle = {Proceedings of the Eleventh Conference on Machine Translation},
  year = {2026},
  address = {Budapest, Hungary},
  publisher = {Association for Computational Linguistics}
}

Contact

The organisers: wmt-qe-metrics-organizers@googlegroups.com. Web page of the shared task: https://www2.statmt.org/wmt26/mteval-task.html.