bert-small-toxicity

This is a toxicity classifier fine-tuned using the gravitee-io/textdetox-multilingual-toxicity-dataset. The model supports a wide range of languages and is trained for toxicity classification ("not-toxic", "toxic").

We perform an 85/15 train-test split per language based on the textdetox dataset. All credits go to the authors of the original corpora.

Performance Overview

While the model performance differs from gravitee-io/distilbert-multilingual-toxicity-classifier Some languages still make the cut, even with the base model being lightweight and trained on English as per the model card.

Original model

Language	eval F1	train F1	Δ F1
en	0.962567	0.992951	-0.030384
fr	0.907895	0.988053	-0.080159
ru	0.904891	0.939517	-0.034626
hi	0.887978	0.942063	-0.054085
de	0.886792	0.972123	-0.085330
uk	0.880000	0.929799	-0.049799
tt	0.836763	0.898663	-0.061901
it	0.824742	0.940903	-0.116160
es	0.817708	0.941259	-0.123551
ja	0.730458	0.795933	-0.065475
hin	0.723944	0.867925	-0.143981
ar	0.688396	0.755972	-0.067576
am	0.626697	0.679577	-0.052881
he	0.570093	0.680567	-0.110474
zh	0.615169	0.648622	-0.033454

Quantized model (ONNX)

Language	eval F1	train F1	Δ F1
en	0.960864	0.993184	-0.032321
fr	0.911958	0.988037	-0.076079
ru	0.895890	0.938834	-0.042944
hi	0.886486	0.939657	-0.053171
de	0.882038	0.970994	-0.088956
uk	0.879892	0.924596	-0.044704
tt	0.828532	0.898537	-0.070004
it	0.826255	0.937281	-0.111027
es	0.821990	0.940571	-0.118581
ja	0.716459	0.791557	-0.075098
hin	0.718750	0.866036	-0.147286
ar	0.671916	0.752080	-0.080164
am	0.630045	0.681464	-0.051419
he	0.563107	0.680237	-0.117131
zh	0.610795	0.635040	-0.024245

🤗 Usage

from transformers import AutoTokenizer
from optimum.onnxruntime import ORTModelForSequenceClassification
import numpy as np
# Load model and tokenizer using optimum
model = ORTModelForSequenceClassification.from_pretrained(
 "gravitee-io/bert-small-toxicity",
 file_name="model.quant.onnx"
)
tokenizer = AutoTokenizer.from_pretrained("gravitee-io/bert-small-toxicity")
# Tokenize input
text = "Your text here"
inputs = tokenizer(text, return_tensors="pt", padding=True, truncation=True)
# Run inference
outputs = model(**inputs)
logits = outputs.logits
# Optional: convert to probabilities
probs = 1 / (1 + np.exp(-logits))
print(probs)

Github Repository

You can check details on how the model was fine-tuned and evaluated on the Github Repository

License

This model is licensed under OpenRAIL++

Citation

@misc{bhargava2021generalization,
      title={Generalization in NLI: Ways (Not) To Go Beyond Simple Heuristics}, 
      author={Prajjwal Bhargava and Aleksandr Drozd and Anna Rogers},
      year={2021},
      eprint={2110.01518},
      archivePrefix={arXiv},
      primaryClass={cs.CL}
}

@article{DBLP:journals/corr/abs-1908-08962,
  author    = {Iulia Turc and
               Ming{-}Wei Chang and
               Kenton Lee and
               Kristina Toutanova},
  title     = {Well-Read Students Learn Better: The Impact of Student Initialization
               on Knowledge Distillation},
  journal   = {CoRR},
  volume    = {abs/1908.08962},
  year      = {2019},
  url       = {http://arxiv.org/abs/1908.08962},
  eprinttype = {arXiv},
  eprint    = {1908.08962},
  timestamp = {Thu, 29 Aug 2019 16:32:34 +0200},
  biburl    = {https://dblp.org/rec/journals/corr/abs-1908-08962.bib},
  bibsource = {dblp computer science bibliography, https://dblp.org}
}

@inproceedings{dementieva2024overview,
  title={Overview of the Multilingual Text Detoxification Task at PAN 2024},
  author={Dementieva, Daryna and Moskovskiy, Daniil and Babakov, Nikolay and Ayele, Abinew Ali and Rizwan, Naquee and Schneider, Frolian and Wang, Xintog and Yimam, Seid Muhie and Ustalov, Dmitry and Stakovskii, Elisei and Smirnova, Alisa and Elnagar, Ashraf and Mukherjee, Animesh and Panchenko, Alexander},
  booktitle={Working Notes of CLEF 2024 - Conference and Labs of the Evaluation Forum},
  editor={Guglielmo Faggioli and Nicola Ferro and Petra Galu{{s}}{{c}}{'a}kov{'a} and Alba Garc{'i}a Seco de Herrera},
  year={2024},
  organization={CEUR-WS.org}
}
@inproceedings{dementieva
-etal-2024-toxicity,
  title = "Toxicity Classification in {U}krainian",
  author = "Dementieva, Daryna and Khylenko, Valeriia and Babakov, Nikolay and Groh, Georg",
  booktitle = "Proceedings of the 8th Workshop on Online Abuse and Harms (WOAH 2024)",
  month = jun,
  year = "2024",
  address = "Mexico City, Mexico",
  publisher = "Association for Computational Linguistics",
  url = "https://aclanthology.org/2024.woah-1.19/",
  doi = "10.18653/v1/2024.woah-1.19",
  pages = "244--255"
}
@inproceedings{DBLP:conf/ecir/BevendorffCCDEFFKMMPPRRSSSTUWZ24,
  author = {Janek Bevendorff and et al.},
  title = {Overview of {PAN} 2024: Multi-author Writing Style Analysis, Multilingual Text Detoxification, Oppositional Thinking Analysis, and Generative {AI} Authorship Verification - Extended Abstract},
  booktitle = {ECIR 2024, Glasgow, UK, March 24-28, 2024, Proceedings, Part {VI}},
  series = {Lecture Notes in Computer Science},
  volume = {14613},
  pages = {3--10},
  publisher = {Springer},
  year = {2024},
  doi = {10.1007/978-3-031-56072-9_1}
}