vietnamese-toxicity

A Vietnamese toxicity classifier fine-tuned from uitnlp/visobert (a Vietnamese-native, social-media pretrained backbone). Single output label toxicity (binary, sigmoid).

v2 (current): fine-tuned from ViSoBERT (Vietnamese-native backbone). Outperforms v1 (fine-tuned from a general multilingual toxicity model) on every threshold — see evaluation below.

Usage

from transformers import pipeline

pipe = pipeline("text-classification", model="dangvansam/vietnamese-toxicity")
pipe("mày là thằng ngu")
# [{'label': 'toxicity', 'score': 0.99...}]

ONNX export available for faster CPU inference:

from optimum.onnxruntime import ORTModelForSequenceClassification
from transformers import AutoTokenizer

tok = AutoTokenizer.from_pretrained("dangvansam/vietnamese-toxicity")
model = ORTModelForSequenceClassification.from_pretrained("dangvansam/vietnamese-toxicity")

Training data

• ViHSD — ~24k Vietnamese social-media comments, 3-class (clean/offensive/hate) collapsed to binary.
• ViCTSD — ~10k Vietnamese news-comment pairs with a binary toxicity label.
• A small set of clean, non-toxic domain questions added as hard negatives to reduce false positives on frustrated-but-non-toxic customer messages.

Combined and deduplicated: 34k training examples (5.4k toxic / ~28.8k clean).

Evaluation

Held-out test splits (never seen during training), combined threshold sweep:

Threshold	Recall	Precision	FP rate
0.3	0.86	0.56	0.13
0.5	0.81	0.62	0.10
0.7	0.74	0.68	0.07
0.9	0.63	0.78	0.03

Comparison against a v1 checkpoint (identical data/recipe, but fine-tuned from a general multilingual toxicity model instead of a Vietnamese-native backbone) at the same thresholds:

Threshold	v1 recall/precision/FP	v2 recall/precision/FP
0.5	0.70 / 0.52 / 0.13	0.81 / 0.62 / 0.10
0.7	0.60 / 0.60 / 0.08	0.74 / 0.68 / 0.07
0.9	0.43 / 0.73 / 0.03	0.63 / 0.78 / 0.03

Also compared against other publicly available Vietnamese hate-speech classifiers on the same held-out test data — this model matches or exceeds them on both source datasets, without any detectable train/test overlap (verified per-source, and against a small hand-written zero-leak probe set with unambiguous VN toxic/clean examples, on which this model scores 20/20).

Label schema

Output label is named toxicity (single label) — sufficient for detectors that take the max score across a recognized-label allowlist rather than requiring a full multi-label taxonomy.