evalstone/evalscope/docs/en/benchmarks/wmt24pp.md

WMT2024++

Overview

WMT2024++ is a comprehensive machine translation benchmark based on the WMT 2024 news translation task. It supports 54 language pairs with English as the source language, enabling evaluation of translation quality across diverse target languages.

Task Description

• Task Type: Machine Translation
• Input: Source text in English with translation prompt
• Output: Translated text in the target language
• Language Pairs: 54 pairs (English to 54 target languages)

Key Features

• Extensive multilingual coverage (54 target languages)
• News domain text for real-world applicability
• Multiple evaluation metrics (BLEU, BERTScore, COMET)
• Standardized prompt template for consistent evaluation
• Supports batch scoring for efficiency

Evaluation Notes

• Default configuration uses 0-shot evaluation
• Metrics: BLEU, BERTScore (XLM-RoBERTa), COMET (wmt22-comet-da)
• Evaluates on test split
• Language-specific normalization applied
• COMET metric requires unbabel-comet package
• Subsets represent individual language pairs (e.g., en-zh_cn, en-de_de)

Properties

Property	Value
Benchmark Name	wmt24pp
Dataset ID	extraordinarylab/wmt24pp
Paper	N/A
Tags	MachineTranslation, MultiLingual
Metrics	bleu, bert_score, comet
Default Shots	0-shot
Evaluation Split	test

Data Statistics

Metric	Value
Total Samples	52,800
Prompt Length (Mean)	265.45 chars
Prompt Length (Min/Max)	71 / 1047 chars

Per-Subset Statistics:

Subset	Samples	Prompt Mean	Prompt Min	Prompt Max
en-ar_eg	960	263.26	75	1039
en-ar_sa	960	263.26	75	1039
en-bg_bg	960	269.26	81	1045
en-bn_in	960	265.26	77	1041
en-ca_es	960	265.26	77	1041
en-cs_cz	960	261.26	73	1037
en-da_dk	960	263.26	75	1039
en-de_de	960	263.26	75	1039
en-el_gr	960	261.26	73	1037
en-es_mx	960	265.26	77	1041
en-et_ee	960	267.26	79	1043
en-fa_ir	960	261.26	73	1037
en-fi_fi	960	265.26	77	1041
en-fil_ph	960	267.26	79	1043
en-fr_ca	960	263.26	75	1039
en-fr_fr	960	263.26	75	1039
en-gu_in	960	267.26	79	1043
en-he_il	960	263.26	75	1039
en-hi_in	960	261.26	73	1037
en-hr_hr	960	267.26	79	1043
en-hu_hu	960	269.26	81	1045
en-id_id	960	271.26	83	1047
en-is_is	960	269.26	81	1045
en-it_it	960	265.26	77	1041
en-ja_jp	960	267.26	79	1043
en-kn_in	960	265.26	77	1041
en-ko_kr	960	263.26	75	1039
en-lt_lt	960	271.26	83	1047
en-lv_lv	960	265.26	77	1041
en-ml_in	960	269.26	81	1045
en-mr_in	960	265.26	77	1041
en-nl_nl	960	261.26	73	1037
en-no_no	960	269.26	81	1045
en-pa_in	960	265.26	77	1041
en-pl_pl	960	263.26	75	1039
en-pt_br	960	271.26	83	1047
en-pt_pt	960	271.26	83	1047
en-ro_ro	960	267.26	79	1043
en-ru_ru	960	265.26	77	1041
en-sk_sk	960	263.26	75	1039
en-sl_si	960	269.26	81	1045
en-sr_rs	960	265.26	77	1041
en-sv_se	960	265.26	77	1041
en-sw_ke	960	265.26	77	1041
en-sw_tz	960	265.26	77	1041
en-ta_in	960	261.26	73	1037
en-te_in	960	263.26	75	1039
en-th_th	960	259.26	71	1035
en-tr_tr	960	265.26	77	1041
en-uk_ua	960	269.26	81	1045
en-ur_pk	960	259.26	71	1035
en-vi_vn	960	271.26	83	1047
en-zh_cn	960	267.26	79	1043
en-zh_tw	960	267.26	79	1043
en-zu_za	960	259.26	71	1035

Sample Example

Subset: en-ar_eg

{
  "input": [
    {
      "id": "557f3aa1",
      "content": [
        {
          "text": "Translate the following english sentence into arabic:\n\nenglish: Siso's depictions of land, water center new gallery exhibition\narabic:"
        }
      ]
    }
  ],
  "target": "رسومات سيسو عن الأرض والمية في معرضه الجديد",
  "id": 0,
  "group_id": 0,
  "subset_key": "en-ar_eg",
  "metadata": {
    "source_text": "Siso's depictions of land, water center new gallery exhibition",
    "target_text": "رسومات سيسو عن الأرض والمية في معرضه الجديد",
    "source_language": "en",
    "target_language": "ar_eg"
  }
}

Prompt Template

Prompt Template:

Translate the following {source_language} sentence into {target_language}:

{source_language}: {source_text}
{target_language}:

Usage

Using CLI

evalscope eval \
    --model YOUR_MODEL \
    --api-url OPENAI_API_COMPAT_URL \
    --api-key EMPTY_TOKEN \
    --datasets wmt24pp \
    --limit 10  # Remove this line for formal evaluation

Using Python

from evalscope import run_task
from evalscope.config import TaskConfig

task_cfg = TaskConfig(
    model='YOUR_MODEL',
    api_url='OPENAI_API_COMPAT_URL',
    api_key='EMPTY_TOKEN',
    datasets=['wmt24pp'],
    dataset_args={
        'wmt24pp': {
            # subset_list: ['en-ar_eg', 'en-ar_sa', 'en-bg_bg']  # optional, evaluate specific subsets
        }
    },
    limit=10,  # Remove this line for formal evaluation
)

run_task(task_cfg=task_cfg)