Beyond declarative toxicity (slurs, hate speech, threats in the words), the classifier is trained to flag politely-phrased harmful requests: harm solicitation, doxxing/privacy requests, solicited discrimination, atrocity denial, and illicit-capability questions.

Performance

Metric	Score
loss	0.1001
accuracy	0.9834
precision	0.9507
recall	0.8914
f1	0.9201

Optimized threshold: 0.6600 (val recall: 0.8854)

Usage (ONNX)

python

from optimum.onnxruntime import ORTModelForSequenceClassification
from transformers import AutoTokenizer

model = ORTModelForSequenceClassification.from_pretrained(
    "HikmaAI/hikmaai-xlm-roberta-large-toxicity",
    subfolder="onnx/fp16",
)
tokenizer = AutoTokenizer.from_pretrained(
    "HikmaAI/hikmaai-xlm-roberta-large-toxicity",
    subfolder="tokenizer",
)

inputs = tokenizer("You are worthless and everyone hates you", return_tensors="pt")
outputs = model(**inputs)
# outputs.logits -> [benign_score, toxic_score]

Training

Epochs: 5
Learning rate: 2e-05
Batch size: 32
Class weights: [1.0, 1.0]
Dataset: multilingual (11 languages), ~90/10 benign/toxic, 12+ sources + targeted synthetic data (request-style harm cells in all 11 languages)

License

Apache-2.0

Citation

bibtex

@misc{hikmaai-toxicity-2026,
  title={hikmaai-xlm-roberta-large-toxicity},
  author={HikmaAI},
  year={2026},
  publisher={HuggingFace},
  url={https://huggingface.co/HikmaAI/hikmaai-xlm-roberta-large-toxicity}
}