Dataset Viewer
Auto-converted to Parquet Duplicate
The dataset viewer is not available for this split.
Couldn't cast array of type struct<id: list<element: string>, sentence: list<element: string>, gold_label: list<element: string>, labels: list<element: struct<human_id: list<element: string>, label: list<element: string>>>> to List({'id': Value('string'), 'sentence': Value('string'), 'gold_label': Value('string'), 'labels': List({'human_id': Value('string'), 'label': Value('string')})})
Error code:   UnexpectedError

Need help to make the dataset viewer work? Make sure to review how to configure the dataset viewer, and open a discussion for direct support.

StereoSet-PT

Dataset Description

StereoSet-PT is a Portuguese-language adaptation of the StereoSet benchmark for measuring stereotypical bias in language models. The dataset was translated from English to Brazilian Portuguese using a pipeline combining automatic translation (Claude Sonnet 4.5) with extensive manual curation by three annotators.

The adaptation covers a stratified 25% sample of the original StereoSet development split, preserving bias domains of gender, profession, race, and religion in both intrasentence and intersentence tasks.

  • Paper: StereoSet-PT: The stereotype evaluation framework for Brazilian Portuguese language models (BRACIS 2026)
  • Original dataset: McGill-NLP/stereoset
  • License: CC-BY-SA 4.0

Supported Tasks

Task Description
Intrasentence Cloze-based task: a context sentence with a BLANK token is completed with one of three options (stereotype, anti-stereotype, or unrelated).
Intersentence Discourse-level task: a context sentence is followed by a candidate sentence that can be stereotypical, anti-stereotypical, or unrelated.

Dataset Structure

The dataset follows the same schema as the original StereoSet. Each instance contains:

  • id: unique identifier
  • target: the social group being evaluated (e.g., enfermeiro, muçulmano)
  • bias_type: one of gender, profession, race, religion
  • context: the context sentence (with BLANK for intrasentence, or a standalone sentence for intersentence)
  • sentences: list of three candidate sentences, each with a gold_label:
    • stereotype
    • anti-stereotype
    • unrelated

Evaluation Metrics

This dataset uses the three metrics defined by the original StereoSet framework (Nadeem et al., 2021):

Stereotype Score (SS)

Measures the proportion of instances where the model assigns a higher probability to the stereotypical sentence over the anti-stereotypical one. The ideal unbiased value is 50.0.

SS=100DxD1(Sstereo(x)>Santi(x))SS = \frac{100}{|D|} \sum_{x \in D} \mathbb{1}(S^{(x)}_{\text{stereo}} > S^{(x)}_{\text{anti}})

Language Modeling Score (LMS)

Measures the proportion of instances where a meaningful sentence (stereotypical or anti-stereotypical) receives a higher score than the unrelated sentence. Higher is better; the ideal value is 100.

LMS=100DxD1(max(Sstereo(x),Santi(x))>Sunrelated(x))LMS = \frac{100}{|D|} \sum_{x \in D} \mathbb{1}(\max(S^{(x)}_{\text{stereo}}, S^{(x)}_{\text{anti}}) > S^{(x)}_{\text{unrelated}})

Idealized Context Association Test (ICAT)

The composite metric that rewards high linguistic competence combined with low bias. The optimal value is 100.

ICAT=LMS×min(SS,  100SS)50ICAT = LMS \times \frac{\min(SS,\; 100 - SS)}{50}


Benchmark Results

Results obtained on the curated Portuguese sample. For reference, the English BERT-base baseline corresponds to the full StereoSet development split as reported by Nadeem et al. (2021).

Baseline Validation (Table 2)

Model Dataset / Language LMS (↑) SS (≈ 50) ICAT (↑)
BERT-base (EN) StereoSet Original (EN) 85.76 59.56 69.37
mBERT Base Curated Sample (PT) 74.06 57.49 62.97
BERTimbau Base Curated Sample (PT) 85.04 56.70 73.64

Higher is better for LMS and ICAT (↑). The ideal Stereotype Score (SS) is 50.

Cross-lingual Comparison — LLMs (Table 3)

Model LMS (EN) SS (EN) ICAT (EN) LMS (PT) SS (PT) ICAT (PT)
Gemma 3 27B 80.80 64.31 57.68 69.35 55.10 62.27
Llama 3.1 8B 80.23 62.87 59.57 63.74 56.32 55.68
Qwen 3.5 27B 81.85 62.33 61.67 67.30 58.01 56.52
Qwen 2.5 1.5B 81.66 61.04 63.64 66.34 57.21 56.78
Sabiá 7B 67.59 58.27 56.42 60.22 56.77 52.07

Encoders vs. Decoders — Curated Sample (Table 4)

Model Type LMS (↑) SS (≈ 50) ICAT (↑)
BERTimbau Base PT-BR Encoder 85.04 56.70 73.64
BERTimbau Large PT-BR Encoder 84.78 58.25 70.80
mBERT Base Multilingual Encoder 74.06 57.49 62.97
Gemma 3 27B Multilingual Decoder 69.35 55.10 62.27
Qwen 3.5 27B Multilingual Decoder 67.30 58.01 56.52
Qwen 2.5 1.5B Multilingual Decoder 66.34 57.21 56.78
Llama 3.1 8B Multilingual Decoder 63.74 56.32 55.68
Sabiá 7B PT-BR Decoder 60.22 56.77 52.07

Dataset Creation

Source Data

A stratified random sample (25%) of the original StereoSet development split was extracted, preserving proportional representation across bias domains. The original development split contains 2,106 intrasentence and 2,123 intersentence instances.

Translation Pipeline

  1. Automatic translation: Claude Sonnet 4.5 (Anthropic) via AWS infrastructure, with temperature 0.1 and top-p 0.9 to minimize hallucinations.
  2. Manual curation: All instances were reviewed and corrected by three annotators using a side-by-side tabular interface displaying the original English and the translated Portuguese texts.

Limitations

  • Cultural asymmetry: The original StereoSet was designed to reflect U.S. social dynamics. Some stereotypes — particularly around race/ethnicity (e.g., Latino immigration tropes) — do not translate directly to the Brazilian context.
  • Missing Brazilian axes: The dataset does not cover discrimination axes specific to Brazil, such as regional xenophobia (North/Northeast prejudice), intolerance toward Afro-Brazilian religions, or the intersection of race, colorism, and social class.
  • Partial coverage: This release corresponds to 25% of the original dataset. Full coverage with native Brazilian stereotypes is proposed as future work.

Citation

If you use this dataset, please cite the following paper:

@inproceedings{aleixo2024stereosetpt,
  title     = {{StereoSet-PT}: The stereotype evaluation framework for {Brazilian Portuguese} language models},
  author    = {Aleixo, Diego Santana and
               Costa, Rosimeire Pereira da and
               Pellicer, Lucas Francisco Amaral Orosco and
               Almeida, Renato Augusto},
  booktitle = {Brazilian Conference on Intelligent Systems (BRACIS)},
  year      = {2024},
  institution = {Instituto de Ci\^{e}ncia e Tecnologia Ita\'{u} (ICTi), S\~{a}o Paulo, Brazil}
}

Also cite the original StereoSet:

@inproceedings{nadeem-etal-2021-stereoset,
  title     = {{S}tereo{S}et: Measuring stereotypical bias in pretrained language models},
  author    = {Nadeem, Moin and Bethke, Anna and Reddy, Siva},
  booktitle = {Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics (ACL)},
  pages     = {5356--5371},
  year      = {2021}
}

Terms of Use

The automatic translation step of this dataset was produced using Claude Sonnet 4.5 (Anthropic) via AWS infrastructure. Use of Claude is governed by Anthropic's Usage Policy and Commercial Terms of Service. Users and researchers who build upon or redistribute this dataset should be aware that the translation outputs were generated with an AI model subject to those terms.


License

This dataset is released under CC-BY-SA 4.0, inheriting the license of the original StereoSet dataset.

All data comply with the Brazilian General Data Protection Law (Lei Geral de Proteção de Dados — LGPD).


Acknowledgments

This research was financially supported by the Instituto de Ciência e Tecnologia Itaú (ICTi), São Paulo, Brazil. The opinions expressed are solely those of the authors and do not reflect the views of Itaú Unibanco or ICTi.


Note: This document is available both in English and Portuguese. The Portuguese version can be found below.


StereoSet-PT

Descrição do Dataset

O StereoSet-PT é uma adaptação para o português brasileiro do benchmark StereoSet, desenvolvido para medir viés estereotipado em modelos de linguagem. O dataset foi traduzido do inglês para o português brasileiro por meio de um pipeline combinando tradução automática (Claude Sonnet) com curadoria manual extensiva realizada por três anotadores.

A adaptação cobre uma amostra estratificada de 25% do split de desenvolvimento original do StereoSet, preservando a representação proporcional dos domínios de viés de gênero, profissão, raça e religião, nas tarefas de intrasentence e intersentence.

  • Artigo: StereoSet-PT: The stereotype evaluation framework for Brazilian Portuguese language models (BRACIS 2026)
  • Dataset original: McGill-NLP/stereoset
  • Licença: CC-BY-SA 4.0

Tarefas Suportadas

Tarefa Descrição
Intrasentence Tarefa de cloze: uma frase de contexto com o token BLANK é completada com uma das três opções (estereótipo, anti-estereótipo ou não relacionada).
Intersentence Tarefa de discurso: uma frase de contexto é seguida por uma frase candidata que pode ser estereotipada, anti-estereotipada ou não relacionada.

Estrutura do Dataset

O dataset segue o mesmo esquema do StereoSet original. Cada instância contém:

  • id: identificador único
  • target: o grupo social avaliado (ex.: enfermeiro, muçulmano)
  • bias_type: um dos valores gender, profession, race, religion
  • context: a frase de contexto (com BLANK para intrasentence, ou frase independente para intersentence)
  • sentences: lista de três frases candidatas, cada uma com um gold_label:
    • stereotype
    • anti-stereotype
    • unrelated

Métricas de Avaliação

Este dataset utiliza as três métricas definidas pelo framework StereoSet original (Nadeem et al., 2021):

Stereotype Score (SS)

Mede a proporção de instâncias em que o modelo atribui maior probabilidade à frase estereotipada em relação à anti-estereotipada. O valor ideal para um modelo sem viés é 50,0.

SS=100DxD1(Sstereo(x)>Santi(x))SS = \frac{100}{|D|} \sum_{x \in D} \mathbb{1}(S^{(x)}_{\text{stereo}} > S^{(x)}_{\text{anti}})

Language Modeling Score (LMS)

Mede a proporção de instâncias em que uma frase com sentido (estereotipada ou anti-estereotipada) recebe pontuação maior do que a frase não relacionada. Quanto maior, melhor; o valor ideal é 100.

LMS=100DxD1(max(Sstereo(x),Santi(x))>Sunrelated(x))LMS = \frac{100}{|D|} \sum_{x \in D} \mathbb{1}(\max(S^{(x)}_{\text{stereo}}, S^{(x)}_{\text{anti}}) > S^{(x)}_{\text{unrelated}})

Idealized Context Association Test (ICAT)

Métrica composta que recompensa alta competência linguística combinada com baixo viés. O valor ideal é 100.

ICAT=LMS×min(SS,  100SS)50ICAT = LMS \times \frac{\min(SS,\; 100 - SS)}{50}


Resultados do Benchmark

Resultados obtidos na amostra curada em português. Para referência, o baseline BERT-base em inglês corresponde ao split de desenvolvimento completo do StereoSet, conforme reportado por Nadeem et al. (2021).

Validação do Baseline (Tabela 2)

Modelo Dataset / Idioma LMS (↑) SS (≈ 50) ICAT (↑)
BERT-base (EN) StereoSet Original (EN) 85.76 59.56 69.37
mBERT Base Amostra Curada (PT) 74.06 57.49 62.97
BERTimbau Base Amostra Curada (PT) 85.04 56.70 73.64

LMS e ICAT: quanto maior. melhor (↑). O valor ideal do Stereotype Score (SS) é 50.

Comparação Multilíngue — LLMs (Tabela 3)

Modelo LMS (EN) SS (EN) ICAT (EN) LMS (PT) SS (PT) ICAT (PT)
Gemma 3 27B 80.80 64.31 57.68 69.35 55.10 62.27
Llama 3.1 8B 80.23 62.87 59.57 63.74 56.32 55.68
Qwen 3.5 27B 81.85 62.33 61.67 67.30 58.01 56.52
Qwen 2.5 1.5B 81.66 61.04 63.64 66.34 57.21 56.78
Sabiá 7B 67.59 58.27 56.42 60.22 56.77 52.07

Encoders vs. Decoders — Amostra Curada (Tabela 4)

Modelo Tipo LMS (↑) SS (≈ 50) ICAT (↑)
BERTimbau Base Encoder PT-BR 85.04 56.70 73.64
BERTimbau Large Encoder PT-BR 84.78 58.25 70.80
mBERT Base Encoder Multilíngue 74.06 57.49 62.97
Gemma 3 27B Decoder Multilíngue 69.35 55.10 62.27
Qwen 3.5 27B Decoder Multilíngue 67.30 58.01 56.52
Qwen 2.5 1.5B Decoder Multilíngue 66.34 57.21 56.78
Llama 3.1 8B Decoder Multilíngue 63.74 56.32 55.68
Sabiá 7B Decoder PT-BR 60.22 56.77 52.07

Criação do Dataset

Dados de Origem

Uma amostra aleatória estratificada (25%) do split de desenvolvimento original do StereoSet foi extraída, preservando a representação proporcional entre os domínios de viés. O split de desenvolvimento original contém 2.106 instâncias intrasentence e 2.123 instâncias intersentence.

Pipeline de Tradução

  1. Tradução automática: Claude Sonnet 4.5 (Anthropic) via infraestrutura AWS, com temperatura 0,1 e top-p 0,9 para minimizar alucinações.
  2. Curadoria manual: Todas as instâncias foram revisadas e corrigidas por três anotadores por meio de uma interface tabular lado a lado exibindo o texto original em inglês e a tradução em português.

Limitações

  • Assimetria cultural: O StereoSet original foi projetado para refletir dinâmicas sociais dos EUA. Alguns estereótipos — especialmente relacionados a raça/etnia (ex.: tropes sobre imigração latina) — não se traduzem diretamente para o contexto brasileiro.
  • Eixos brasileiros ausentes: O dataset não cobre eixos de discriminação específicos do Brasil, como xenofobia regional (preconceito contra Norte/Nordeste), intolerância às religiões afro-brasileiras, ou a interseção de raça, colorismo e classe social.
  • Cobertura parcial: Esta versão corresponde a 25% do dataset original. Cobertura completa com estereótipos tipicamente brasileiros é proposta como trabalho futuro.

Citação

Se utilizar este dataset, por favor cite o seguinte artigo:

@inproceedings{aleixo2024stereosetpt,
  title     = {{StereoSet-PT}: The stereotype evaluation framework for {Brazilian Portuguese} language models},
  author    = {Aleixo, Diego Santana and
               Costa, Rosimeire Pereira da and
               Pellicer, Lucas Francisco Amaral Orosco and
               Almeida, Renato Augusto},
  booktitle = {Brazilian Conference on Intelligent Systems (BRACIS)},
  year      = {2024},
  institution = {Instituto de Ci\^{e}ncia e Tecnologia Ita\'{u} (ICTi), S\~{a}o Paulo, Brazil}
}

Cite também o StereoSet original:

@inproceedings{nadeem-etal-2021-stereoset,
  title     = {{S}tereo{S}et: Measuring stereotypical bias in pretrained language models},
  author    = {Nadeem, Moin and Bethke, Anna and Reddy, Siva},
  booktitle = {Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics (ACL)},
  pages     = {5356--5371},
  year      = {2021}
}

Termos de Uso

A etapa de tradução automática deste dataset foi realizada com o Claude Sonnet 4.5 (Anthropic) via infraestrutura AWS. O uso do Claude é regido pela Política de Uso Aceitável e pelos Termos de Serviço Comercial da Anthropic. Usuários e pesquisadores que utilizem ou redistribuam este dataset devem estar cientes de que as saídas de tradução foram geradas por um modelo de IA sujeito a esses termos.


Licença

Este dataset é disponibilizado sob a licença CC-BY-SA 4.0, herdando a licença do dataset StereoSet original.

Todos os dados estão em conformidade com a Lei Geral de Proteção de Dados (LGPD).


Agradecimentos

Esta pesquisa foi financiada pelo Instituto de Ciência e Tecnologia Itaú (ICTi), São Paulo, Brasil. As opiniões expressas são exclusivamente dos autores e não refletem as visões do Itaú Unibanco ou do ICTi.

Downloads last month
1