gravitee-io/distilbert-multilingual-toxicity-classifier

distillbert-multilingual-toxicity-classifier

This is a DistilBERT-based multilingual 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

Original model

Language	Validation F1	Training F1	ΔF1
Russian	0.9572	0.9897	-0.0324
English	0.9528	0.9853	-0.0325
Hindi	0.9248	0.9599	-0.0351
Armenian	0.6513	0.6915	-0.0402
French	0.9446	0.9874	-0.0428
Tatar	0.9200	0.9682	-0.0482
Ukrainian	0.8997	0.9511	-0.0514
Japanese	0.8658	0.9253	-0.0595
German	0.8904	0.9547	-0.0643
Spanish	0.8564	0.9399	-0.0835
Chinese	0.6865	0.7807	-0.0942
Arabic	0.7563	0.8550	-0.0987
Italian	0.8223	0.9271	-0.1048
Hinglish	0.7234	0.8533	-0.1299
Hebrew	0.6455	0.8441	-0.1987

Quantized model (ONNX)

Language	Val F1	Quantized Val F1	Δ Val F1	Train F1	Quantized Train F1	Δ Train F1
Russian	0.9572	0.9609	+0.0037	0.9897	0.9875	−0.0022
English	0.9528	0.9495	−0.0033	0.9853	0.9857	+0.0004
German	0.8904	0.8842	−0.0062	0.9547	0.9369	−0.0178
Hindi	0.9248	0.8940	−0.0300	0.9599	0.9335	−0.0264
French	0.9446	0.9351	−0.0095	0.9874	0.9814	−0.0060
Japanese	0.8658	0.8584	−0.0074	0.9253	0.9081	−0.0172
Ukrainian	0.8997	0.8988	−0.0009	0.9511	0.9476	−0.0035
Tatar	0.9200	0.9148	−0.0052	0.9682	0.9631	−0.0051
Amharic	0.6513	0.6377	−0.0136	0.6915	0.6863	−0.0052
Spanish	0.8564	0.8439	−0.0125	0.9399	0.9273	−0.0126
Chinese	0.6865	0.6697	−0.0168	0.7807	0.7596	−0.0211
Arabic	0.7563	0.7535	−0.0028	0.8550	0.8481	−0.0069
Italian	0.8223	0.8033	−0.0190	0.9271	0.9193	−0.0078
Hinglish	0.7234	0.7260	+0.0026	0.8533	0.8436	−0.0097
Hebrew	0.6455	0.6190	−0.0265	0.8441	0.8204	−0.0237

The quantized model maintains high accuracy and generalization on the evaluation set, with negligible performance loss for most languages.

🤗 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/distilbert-multilingual-toxicity-classifier",
 file_name="model.quant.onnx"
)
tokenizer = AutoTokenizer.from_pretrained("gravitee-io/distilbert-multilingual-toxicity-classifier")

# 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

@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}
}