"""mDeBERTa-v3 fine-tuning wrapper - multilingual transformer with disentangled attention.

Subclasses :class:`IndoBERTModel` (same architecture: CLS pooling + linear head),
overriding the default ``model_name`` to ``microsoft/mdeberta-v3-base``.
DeBERTa-v3's disentangled attention + ELECTRA-style RTD pretraining were the
top performer on Malay X cyberbullying detection in Singh & Othman (2025,
IJIC 15(1)) - a closely related task to Indonesian TikTok comments.

Cite: He et al. 2021, "DeBERTaV3: Improving DeBERTa using ELECTRA-Style Pre-Training
with Gradient-Disentangled Embedding Sharing".
"""

from __future__ import annotations

from src.models.indobert import IndoBERTModel


class MDeBERTaModel(IndoBERTModel):
    """mDeBERTa-v3 base fine-tuning wrapper. Same interface as IndoBERTModel."""

    def __init__(
        self,
        model_name: str = "microsoft/mdeberta-v3-base",
        **kwargs,
    ) -> None:
        super().__init__(model_name=model_name, **kwargs)