"Hate speech dataset with style perturbation"

"We present a style-perturbed benchmark for robustness evaluation in hate speech detection. The benchmark is constructed from three widely used datasets, namely IHC, SBIC, and DynaHate, and introduces multiple stylistic perturbation settings while preserving the original label semantics as much as possible. For each dataset, we retain the undisturbed version together with several perturbed variants, enabling controlled comparison between clean and style-shifted inputs. The resulting resource is designed to test whether a model relies on task-relevant semantic content or exhibits residual dependence on superficial stylistic cues. In addition to the perturbed texts and class labels, the benchmark preserves sample-level metadata to support reproducible evaluation and fine-grained analysis across perturbation types and annotation rounds. This benchmark provides a practical testbed for studying robustness, prediction consistency, and label stability under semantically preserving stylistic variation, and is particularly suitable for evaluating large language model based classifiers and causal content-style disentanglement methods."