Creating evaluation resources for low-resource languages requires substantial time and scarce linguistic expertise. To preserve SMUGRI-COPA as a useful benchmark, we kindly request that it not be used for model training or reposted in locations accessible to web crawlers.

SMUGRI-COPA

SMUGRI-COPA is a manually translated commonsense causal reasoning benchmark for Võro and Livonian, two heavily under-resourced Finnic languages. It extends COPA and follows the XCOPA structure, enabling comparison with existing XCOPA translations.

For each language, the dataset contains a 500-example test set and a 100-example validation set. The XCOPA split assignments, labels, and option order are preserved.

Access and benchmark preservation

SMUGRI-COPA is released under the Creative Commons Attribution 4.0 International license. The automatically approved gate adds no legal restrictions.

⚠️ Creating evaluation resources for low-resource languages requires substantial time and scarce linguistic expertise. To preserve SMUGRI-COPA as a useful benchmark, please do not use it for model training or repost it where web crawlers can access it. When evaluating through hosted APIs, keep answer labels out of submitted prompts. These requests are non-binding and do not modify the CC BY 4.0 license.

Loading the dataset

Because SMUGRI-COPA is gated, first request access through this repository page and authenticate with Hugging Face. The default revised dataset can then be loaded as follows:

from datasets import load_dataset

repo_id = "tartuNLP/smugri-copa"

datasets = {
    "vro_Latn": load_dataset(repo_id, "vro_Latn"),
    "liv_Latn": load_dataset(repo_id, "liv_Latn"),
}

Versions and reproducibility

Changelog

Revised version with metadata (08.2026): Relative to the paper version, 22 Livonian examples were revised: 17/500 test examples (3.4%) and 5/100 validation examples (5%). Fifteen revisions fix minor typographical errors and seven refine lexical choice; none alter the labels or intended causal relationships. The Võro benchmark content is unchanged. Both language configurations gain item-level ambiguity flags and review comments under metadata.
Paper evaluation version (05.2026): Dataset snapshot used for all results reported in the paper.

Reproducing the paper results

To reproduce the paper results, load the pinned revision:

from datasets import load_dataset

repo_id = "tartuNLP/smugri-copa"
paper_revision = "858805289bd02708703242bdc37eaac68c7ad929"

paper_datasets = {
    "vro_Latn": load_dataset(
        repo_id,
        "vro_Latn",
        revision=paper_revision,
    ),
    "liv_Latn": load_dataset(
        repo_id,
        "liv_Latn",
        revision=paper_revision,
    ),
}

Use the latest version for other evaluations.

Dataset structure

The dataset provides separate Võro and Livonian configurations, each with validation and test splits.

Field	Description
`idx`	Example identifier
`premise`	Description of an everyday situation
`question`	Whether the alternatives describe a plausible cause or effect
`choice1`	First alternative
`choice2`	Second alternative
`label`	Index of the correct alternative
`canary`	Dataset-specific benchmark canary for voluntary filtering and contamination analysis
`metadata.potentially_ambiguous`	Whether the example was identified as potentially ambiguous based on translator comments and subsequent author review
`metadata.comments`	Author-written English explanation of the potential ambiguity based on the review; an empty string when the example was not flagged

Dataset creation

The Võro translation was produced by a native speaker, and the Livonian translation by a fluent speaker. Translators received the English COPA examples and the existing Estonian XCOPA translation. They were asked to preserve the intended causal relationship while producing natural target-language text and could adjust wording when a direct translation was unsuitable.

Unlike the original XCOPA translation procedure, translators were shown the correct labels. This helped them preserve the intended answers and identify examples whose correctness became uncertain after translation.

Intended use

SMUGRI-COPA supports the evaluation of causal commonsense reasoning in Võro and Livonian and comparison across language resource levels. As a discriminative benchmark, it tests selection between alternatives rather than open-ended target-language generation.

ℹ️ SMUGRI-COPA presents one possible translation into each language, not a definitive or authoritative representation. It is neither a comprehensive measure of either language nor evidence of performance across all language understanding or generation tasks.

Limitations

Each language was translated by one person, and independent label-blind human validation was not conducted.
Showing translators the correct labels may have influenced wording choices or introduced unintended surface cues.
Võro and Livonian have multiple language varieties and orthographic practices. This dataset represents one translation rather than the only correct version of either language.
COPA and XCOPA may have appeared in model training data. New translations reduce direct string-level contamination but do not eliminate task-level contamination.
A reasoning model may translate an item into another language and recover an answer encountered through COPA or XCOPA. Accuracy, therefore, cannot by itself establish that the model reasoned independently in the target language.
COPA evaluates one form of causal commonsense reasoning, so results may not generalize to broader language understanding or generation tasks.

License and attribution

SMUGRI-COPA is an adaptation of COPA and XCOPA distributed under CC BY 4.0.

Citation

The SMUGRI-COPA paper citation will be added soon. Please also cite the original COPA and XCOPA papers:

@inproceedings{roemmele2011choice,
    title={Choice of Plausible Alternatives: An Evaluation of Commonsense Causal Reasoning.},
    author={Roemmele, Melissa and Bejan, Cosmin Adrian and Gordon, Andrew S},
    booktitle={AAAI spring symposium: logical formalizations of commonsense reasoning},
    pages={90--95},
    year={2011}
}

@inproceedings{ponti-etal-2020-xcopa,
    title = "{XCOPA}: A Multilingual Dataset for Causal Commonsense Reasoning",
    author = "Ponti, Edoardo Maria  and
      Glava{\v{s}}, Goran  and
      Majewska, Olga  and
      Liu, Qianchu  and
      Vuli{\'c}, Ivan  and
      Korhonen, Anna",
    editor = "Webber, Bonnie  and
      Cohn, Trevor  and
      He, Yulan  and
      Liu, Yang",
    booktitle = "Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP)",
    month = nov,
    year = "2020",
    address = "Online",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2020.emnlp-main.185/",
    doi = "10.18653/v1/2020.emnlp-main.185",
    pages = "2362--2376",
}

Downloads last month: 23

Total file size:

132 kB