DeepPavlov/clinc150_ru

Russian clinc150

This is a text classification dataset. It is intended for machine learning research and experimentation.

This dataset is obtained via formatting another publicly available data to be compatible with our AutoIntent Library.

Usage

It is intended to be used with our AutoIntent Library:

from autointent import Dataset

clinc150_ru = Dataset.from_hub("AutoIntent/clinc150_ru")

Source

This dataset is taken from private github repository LadaNikitina/clinc150 and formatted with our AutoIntent Library:

from autointent import Dataset
from autointent.schemas import Sample
from datasets import load_from_disk, Dataset as HFDataset


def convert_ruclinc150(clinc150_train: HFDataset, ood_index=42):
    all_labels = sorted(clinc150_train.unique("intent"))
    assert all_labels == list(range(151))

    in_domain_samples = clinc150_train.filter(lambda x: x["intent"] != ood_index)
    oos_samples = clinc150_train.filter(lambda x: x["intent"] == ood_index)

    classwise_samples = [[] for _ in range(150)]

    for batch in in_domain_samples.iter(batch_size=16, drop_last_batch=False):
        for txt, intent_id in zip(batch["text"], batch["intent"], strict=False):
            intent_id -= int(intent_id > ood_index)
            target_list = classwise_samples[intent_id]
            target_list.append({"utterance": txt, "label": intent_id})

    train_samples = [sample for samples_from_one_class in classwise_samples for sample in samples_from_one_class]
    oos_samples = [{"utterance": txt} for txt in oos_samples["text"]]

    return [Sample(**sample) for sample in train_samples + oos_samples]

if __name__ == "__main__":
    # git clone git@github.com:LadaNikitina/clinc150 data/RuClinc150
    # rm -rf data/RuClinc150/.git

    clinc150 = load_from_disk("data/RuClinc150")
    train_samples = convert_ruclinc150(clinc150["train"])
    val_samples = convert_ruclinc150(clinc150["validation"])
    test_samples = convert_ruclinc150(clinc150["test"])

    clinc150_converted = Dataset.from_dict(
        {"train": train_samples, "validation": val_samples, "test": test_samples}
    )