Polygl0t/portuguese-toxicity-qwen-annotations

---
dataset_info:
features:
- name: id
dtype: string
- name: text
dtype: string
- name: score
dtype: int64
- name: subset
dtype: string
- name: source
dtype: string
splits:
- name: train
num_bytes: 1296242732
num_examples: 700000
download_size: 795304270
dataset_size: 1296242732
configs:
- config_name: default
data_files:
- split: train
path: data/train-*
license: apache-2.0
task_categories:
- text-classification
language:
- pt
tags:
- toxicity
- portuguese
pretty_name: Portuguese Toxicity Annotations
size_categories:
- 100K<n<1M
---
# Annotations for the Portuguese-Toxicity classifier 📚

## Table of Contents

- [Dataset Description](#dataset-description)
- [Dataset Summary](#dataset-summary)
- [Supported Tasks and Leaderboards](#supported-tasks-and-leaderboards)
- [Languages](#languages)
- [Dataset Structure](#dataset-structure)
- [Data Instances](#data-instances)
- [Data Fields](#data-fields)
- [Subsets and Splits](#subsets-and-splits)
- [Dataset Creation](#dataset-creation)
- [Source Data](#source-data)
- [Annotations](#annotations)
- [Annotation Process](#annotation-process)
- [Considerations for Using the Data](#considerations-for-using-the-data)
- [Social Impact of Dataset](#social-impact-of-dataset)
- [Additional Information](#additional-information)
- [Dataset Maintainers](#dataset-maintainers)
- [Licensing Information](#licensing-information)
- [Citation Information](#citation-information)
- [Acknowledgments](#acknowledgments)
- [Contributions](#contributions)

## Dataset Description

- **Homepage:** https://huggingface.co/datasets/Polygl0t/portuguese-toxicity-qwen-annotations
- **Repository:** https://huggingface.co/datasets/Polygl0t/portuguese-toxicity-qwen-annotations
- **Point of Contact:** [Polyg0t](mailto:[e-mail])

### Dataset Summary

This dataset contains the annotations used for training a toxicity classifier ([Polygl0t/portuguese-bertabaporu-large-toxicity-classifier](https://huggingface.co/Polygl0t/portuguese-bertabaporu-large-toxicity-classifier) and [Polygl0t/portuguese-bertimbau-toxicity-classifier](https://huggingface.co/Polygl0t/portuguese-bertimbau-toxicity-classifier)). These annotations were generated by [Qwen/Qwen2.5-32B-Instruct](https://huggingface.co/Qwen/Qwen2.5-32B-Instruct).

### Supported Tasks and Leaderboards

This dataset can be used for the task of text classification, specifically for toxicity detection in Portuguese text.

### Languages

Portuguese.

## Dataset Structure

### Data Instances

- **id:** a unique identifier for each sample (md5 hash).
- **text:** a string of text in Portuguese.
- **source:** the source where that string originated.
- **subset:** a short string indicating the name of the subset (referring to the original dataset or crawl).
- **score:** the score assigned by [Qwen/Qwen2.5-32B-Instruct](https://huggingface.co/Qwen/Qwen2.5-32B-Instruct).

### Data Fields

```json
{
"text": "Amostra de texto em português para avaliação.",
"source": "https://data.commoncrawl.org/crawl-data/CC-MAIN-2025-30/index.html",
"subset": "CC-MAIN-2025-30",
"id": "a1b2c3d4e5f67890123456789abcdef",
"score": 2
}
```

### Subsets and Splits

This dataset contains a single split: `train`, which includes all 700,000 samples.

```python
from datasets import load_dataset

# Load the main dataset
ds = load_dataset("Polygl0t/portuguese-toxicity-qwen-annotations", split="train")

# If you don't want to download the entire dataset, set streaming to `True`
ds = load_dataset("Polygl0t/portuguese-toxicity-qwen-annotations", split="train", streaming=True)
```

## Dataset Creation

### Source Data

All data was sourced from [Polygl0t/gigaverbo-v2](https://huggingface.co/datasets/Polygl0t/gigaverbo-v2). The **id** field corresponds to the md5 hash of the text.

### Annotations

The dataset was created to provide high-quality toxicity annotations for Portuguese text, leveraging the capabilities of Qwen2.5 to generate consistent and scalable annotations.

The score distribution of this dataset is the following:

| Toxicity Score | N samples |
|----------------|-----------|
| 1 | 605937 |
| 2 | 50381 |
| 3 | 26695 |
| 4 | 11198 |
| 5 | 5789 |

#### Annotation Process

The initial annotation was performed on a random sample of 700,000 documents from [Polygl0t/gigaverbo-v2](https://huggingface.co/datasets/Polygl0t/gigaverbo-v2), where we tried to sample equally from all the subsets that form this corpus. In terms of infrastructure, we used 4 x NVIDIA A100-SXM4-80GB GPUs and the vLLM library as our inference engine with 4-fold tensor parallelism.

Below is the prompt used for annotation:

```text
Você é um filtro de toxicidade para a língua portuguesa. Sua tarefa é avaliar, em uma escala de 1 a 5, o quão tóxico (ou seja, nocivo, ofensivo ou inapropriado) é um determinado trecho de texto. Você DEVE fornecer sua pontuação em formato JSON.

### Critérios de Avaliação:

- Dê 1 se o texto não for tóxico, for neutro e apropriado para todos os públicos. O texto não contém conteúdo ofensivo, prejudicial ou inflamatório.
- Dê 2 se o texto apresentar toxicidade leve, como linguagem levemente inapropriada, leve grosseria ou sentimento negativo sutil, mas sem ser explicitamente ofensivo ou prejudicial.
- Dê 3 se o texto contiver toxicidade moderada, incluindo alguns insultos, linguagem ofensiva ou frases inadequadas, mas sem ser altamente agressivo ou gravemente nocivo.
- Dê 4 se o texto for altamente tóxico, contendo discurso de ódio explícito, ameaças diretas, insultos severos ou estereótipos prejudiciais. Pode também incluir linguagem depreciativa ou agressiva direcionada a indivíduos ou grupos.
- Dê 5 se o texto for extremamente tóxico, exibindo o mais alto nível de conteúdo nocivo, como incitação explícita à violência, discriminação grave, ameaças ou linguagem extremamente ofensiva. Essa pontuação é reservada para os textos mais perigosos e abusivos.

### Requisitos de Saída:

- Sua saída deve ser um objeto JSON válido.
- O JSON deve conter apenas um par chave-valor.
- A chave deve ser 'score'.
- O valor deve ser um número inteiro entre 1 e 5.
- O valor deve representar fielmente os critérios acima.

### Exemplo de Saída:

{'score': 2}

Certifique-se de que suas respostas sigam estritamente este formato e avaliem o texto com precisão com base nos critérios fornecidos.

Abaixo está um trecho retirado de uma página da web. Avalie se o trecho contém linguagem nociva, ofensiva ou inapropriada. Atribua uma pontuação entre 1 e 5. Certifique-se de que sua pontuação reflita o nível de toxicidade presente no texto. O texto será fornecido em português.

Texto: <<O texto a ser avaliado>>

Forneça sua pontuação como um número inteiro entre 1 e 5. Responda apenas em formato JSON.
```

## Considerations for Using the Data

### Social Impact of Dataset

The dataset aims to provide a valuable resource for understanding and mitigating toxic language in Bengali text. By offering high-quality annotations, it can help researchers and developers create more effective moderation tools, promote healthier online interactions, and contribute to the broader field of NLP for low-resource languages. However, it is important to understand that this dataset inevitably contains samples that may be offensive or harmful.

## Additional Information

### Dataset Maintainers

- [Nicholas Kluge Corrêa](mailto:[e-mail]).
- [Shiza Fatimah](mailto:[e-mail]).
- [Aniket Sen](mailto:[e-mail]).

### Licensing Information

The dataset is licensed under the [Apache-2.0 License](LICENSE).

### Citation Information

```latex
@misc{correa2026tucano2cool,
title={{Tucano 2 Cool: Better Open Source LLMs for Portuguese}},
author={Nicholas Kluge Corr{\^e}a and Aniket Sen and Shiza Fatimah and Sophia Falk and Lennard Landgraf and Julia Kastner and Lucie Flek},
year={2026},
eprint={2603.03543},
archivePrefix={arXiv},
primaryClass={cs.CL},
url={https://arxiv.org/abs/2603.03543},
}
```

### Acknowledgments

Polyglot is a project funded by the Federal Ministry of Education and Research (BMBF) and the Ministry of Culture and Science of the State of North Rhine-Westphalia (MWK) as part of TRA Sustainable Futures (University of Bonn) and the Excellence Strategy of the federal and state governments.

We also gratefully acknowledge the granted access to the [Marvin cluster](https://www.hpc.uni-bonn.de/en/systems/marvin) hosted by [University of Bonn](https://www.uni-bonn.de/en) along with the support provided by its High Performance Computing & Analytics Lab.

### Contributions

If you want to contribute, contact us at [[e-mail]](mailto:[e-mail])!