arxiv-pii-join.py

import datasets
import pandas as pd


class ArXivPII(datasets.GeneratorBasedBuilder):
    
    BUILDER_CONFIG_CLASS = datasets.BuilderConfig

    def _info(self) -> datasets.DatasetInfo:
        features = datasets.Features({
            "arxiv_id": datasets.Value("string"),
            "text": datasets.Value("string"),
            "labels": datasets.Sequence({
                "start": datasets.Value("int16"),
                "end": datasets.Value("int16"),
                "label": datasets.ClassLabel(names=["NAME", "EMAIL", "PHONE"]),
            }),
        })
        return datasets.DatasetInfo(
            features=features,
            description="ArXiv PII Named Entity Recognition Dataset",
            version=datasets.Version("1.0.0"),
        )

    def _split_generators(self, dl_manager: datasets.DownloadManager):
        splits = []
        url = "https://huggingface.co/datasets/sorokin/arxiv-pii/raw/main/data/"
        for name in ["2312", "2401"]:
            pages_path = dl_manager.download(url + name + "/pages.csv")
            labels_path = dl_manager.download(url + name + "/labels.csv")
            splits.append(
                datasets.SplitGenerator(
                    name=name,
                    gen_kwargs={"pages_path": pages_path, "labels_path": labels_path},
                )
            )
        return splits

    def _generate_examples(self, pages_path, labels_path):
        example_id = 0
        df_pages = pd.read_csv(pages_path, index_col="arxiv_id")
        df_labels = pd.read_csv(labels_path, index_col="arxiv_id")
        for arxiv_id, page in df_pages.iterrows():
            labels = []
            if arxiv_id in df_labels.index:
                rows = df_labels.loc[[arxiv_id]]
                for _, e in rows.iterrows():
                    labels.append({"start": e.start, "end": e.end, "label": e.label})
            yield example_id, {"arxiv_id": arxiv_id, "text": page.text, "labels": labels}
            example_id += 1