@dataclass
class PreprocessingConfig:
    padding: str = "max_length"
    max_length: int = 512
    tokenizer_name: str = MISSING
    test_size: int = 10000
    train_size: int = 50000


@dataclass
class RobertaPreprocessing(PreprocessingConfig):
    max_length: int = 256
    tokenizer_name: str = "siebert/sentiment-roberta-large-english"
    test_size: int = 5000
    train_size: int = 20000

from hydra.core.config_store import ConfigStore
from src import (
    RobertaPreprocessing,
    DistillBertPreprocessing,
    MainConfig,
)

#...

cs = ConfigStore.instance()
cs.store(group="preprocessing", name="distillbert", node=DistillBertPreprocessing)
cs.store(group="preprocessing", name="roberta", node=RobertaPreprocessing)

@hydra.main(version_base=None, config_name="config")
def main(cfg: MainConfig):
    ...

@dataclass
class ModelConfig:
    _target_: str = "transformers.AutoModelForSequenceClassification.from_pretrained"
    _convert_: str = "object"
    pretrained_model_name_or_path: str = MISSING
    problem_type: str = "single_label_classification"

@hydra.main(version_base=None, config_name="config")
def main(cfg: MainConfig):
    log.info(cfg.preprocessing.pading)

@dataclass
class PreprocessingConfig:
    padding: str = "max_length"
    max_length: int = 512
    tokenizer_name: str = MISSING
    test_size: int = 10000
    train_size: int = 50000


@dataclass
class RobertaPreprocessing(PreprocessingConfig):
    max_length: int = 256
    tokenizer_name: str = "siebert/sentiment-roberta-large-english"
    test_size: int = 5000
    train_size: int = 20000


@dataclass
class DistillBertPreprocessing(PreprocessingConfig):
    tokenizer_name: str = "lxyuan/distilbert-base-multilingual-cased-sentiments-student"


# аналогично для ModelConfig и #TrainingConfig

@dataclass
class MainConfig:
    defaults: list[Any] = field(default_factory=lambda: defaults)
    preprocessing: PreprocessingConfig = MISSING
    model: ModelConfig = MISSING
    training: TrainingConfig = MISSING

defaults = [
    {"preprocessing": "distillbert"},
    {"model": "distillbert"},
    {"training": "distillbert"},
]

@dataclass
class MainConfig:
    defaults: list[Any] = field(default_factory=lambda: defaults)
    preprocessing: PreprocessingConfig = MISSING
    model: ModelConfig = MISSING
    training: TrainingConfig = MISSING

defaults = [
    {"preprocessing": MISSING},
    {"model": "distillbert"},
    {"training": "distillbert"},
]

@dataclass(frozen=True)
class ModelConfig:
    _target_: str = "transformers.AutoModelForSequenceClassification.from_pretrained"
    _convert_: str = "object"
    pretrained_model_name_or_path: str = MISSING
    problem_type: str = "single_label_classification"

@dataclass(frozen=True)
class DistillBertModel(ModelConfig):
    pretrained_model_name_or_path: str = (
        "lxyuan/distilbert-base-multilingual-cased-sentiments-student"
    )
    num_labels: int = 2
    ignore_mismatched_sizes: bool = True