OperKH/twitter-xlmr-toxicity-classifier-ONNX

twitter-xlmr-toxicity-classifier (ONNX)

ONNX export of textdetox/twitter-xlmr-toxicity-classifier, a binary toxicity classifier for 15 languages (Twitter-XLM-RoBERTa large fine-tuned on textdetox/multilingual_toxicity_dataset), for use with Transformers.js and ONNX Runtime. The original repository has PyTorch weights only.

Files

File	Precision	Transformers.js dtype	Size
onnx/model_quantized.onnx	int8 (dynamic, per-channel)	q8	537 MB
onnx/model.onnx + onnx/model.onnx_data	fp32	fp32	2.1 GB

The weights are those of the original model; nothing was retrained. Changes to config.json:

• id2label/label2id were added following the original model card (0 = neutral, 1 = toxic); the original config has no label names
• transformers.js_config.use_external_data_format tells Transformers.js that the fp32 model keeps its weights in a separate .onnx_data file (ONNX protobuf is capped at 2 GB)

Usage

import { pipeline } from '@huggingface/transformers';

const classifier = await pipeline('text-classification', 'OperKH/twitter-xlmr-toxicity-classifier-ONNX', { dtype: 'q8' });

await classifier('You are amazing!', { top_k: null });
// [{ label: 'neutral', score: 0.999 }, { label: 'toxic', score: 0.001 }]

Use dtype: 'fp32' for the unquantized model.

The scores bunch up close to 0 and 1, so useful decision thresholds tend to be near the ends (for example 0.98 rather than 0.5); pick one on your own data.

Conversion

• Export: 🤗 Optimum main_export, task text-classification, opset 17
• Quantization: ONNX Runtime quantize_dynamic, QInt8 weights, per-channel, MatMul nodes only

On a private sample of ~1,600 Russian/Ukrainian chat messages the int8 model scored the same as the fp32 export (ROC-AUC 0.981 for both), at about a quarter of the size and 2.4× the speed on CPU.

License and credits

Same license as the original model: OpenRAIL++, including its use restrictions. All credit for the model goes to its authors, who prepared it for the TextDetox 2025 Shared Task; please cite their work (see the original model card).