ft-xlm-roberta-toxicity

Fine-tuned XLM-RoBERTa for Toxicity Classification in Spanish

Overview

This is a fine-tuned XLM-RoBERTa trained using as a base model XLM-RoBERTa base-sized pre-trained on 2.5TB of filtered CommonCrawl data that comprises 100 languages. The dataset for training this model is a gold standard for protest events for toxicity and incivility in Spanish.

The dataset comprises ~5M data points from three Latin American protest events: (a) protests against the coronavirus and judicial reform measures in Argentina during August 2020; (b) protests against education budget cuts in Brazil in May 2019; and (c) the social outburst in Chile stemming from protests against the underground fare hike in October 2019. We are focusing on interactions in Spanish to elaborate a gold standard for digital interactions in this language, therefore, we prioritise Argentinian and Chilean data.

• GitHub repository
• Dataset on Zenodo
• Reference paper

Labels

• NONTOCIX
• TOXIC

Example Usage

## Pipeline as a high-level helper
from transformers import pipeline
toxic_classifier = pipeline("text-classification", model="bgonzalezbustamante/ft-xlm-roberta-toxicity")

## Non-toxic example
non_toxic = toxic_classifier("Que tengas un excelente día :)")

## Toxic example
toxic = toxic_classifier("Eres un maldito infeliz")

## Print examples
print(non_toxic)
print(toxic)

Output:

[{'label': 'NONTOXIC', 'score': 0.5529471635818481}]
[{'label': 'TOXIC', 'score': 0.6219274401664734}]

Validation Metrics

• Accuracy: 0.740
• Precision: 0.688
• Recall: 0.924
• F1-score: 0.789

Environmental Impact

• 14.9g CO₂ eq emissions

How to Cite

González-Bustamante, B. (2024). ft-xlm-roberta-toxicity (Revision 9f95b0c). Hugging Face. https://doi.org/10.57967/hf/6865.