"""Configuration for the BashkirRoBERTa Pre-LN masked-language model."""

from transformers import PretrainedConfig


class BashkirRobertaConfig(PretrainedConfig):
    """Keeps the exact architecture of the project's flagship checkpoint.

    Despite the familiar name, this is not Hugging Face's post-LayerNorm
    ``RobertaConfig``: the encoder blocks here use pre-LayerNorm.
    """

    model_type = "bashkir-roberta-preln"

    def __init__(
        self,
        vocab_size=16_384,
        hidden_size=640,
        num_hidden_layers=8,
        num_attention_heads=10,
        intermediate_size=2_560,
        max_position_embeddings=256,
        hidden_dropout_prob=0.1,
        pad_token_id=0,
        bos_token_id=2,
        eos_token_id=3,
        cls_token_id=4,
        sep_token_id=5,
        mask_token_id=6,
        tie_word_embeddings=True,
        **kwargs,
    ):
        super().__init__(
            pad_token_id=pad_token_id,
            bos_token_id=bos_token_id,
            eos_token_id=eos_token_id,
            cls_token_id=cls_token_id,
            sep_token_id=sep_token_id,
            mask_token_id=mask_token_id,
            tie_word_embeddings=tie_word_embeddings,
            **kwargs,
        )
        self.vocab_size = vocab_size
        self.hidden_size = hidden_size
        self.num_hidden_layers = num_hidden_layers
        self.num_attention_heads = num_attention_heads
        self.intermediate_size = intermediate_size
        self.max_position_embeddings = max_position_embeddings
        self.hidden_dropout_prob = hidden_dropout_prob