dataset.py

import pandas as pd
from datasets import DatasetInfo, Features, GeneratorBasedBuilder, Sequence, Split, SplitGenerator, Value


class PodcastConversationsWithMetadataAndEmbedding(GeneratorBasedBuilder):
    def _info(self):
        return DatasetInfo(
            features=Features(
                {
                    "c_guid": Value("string"),
                    "participants": Sequence(Value("string")),
                    "transcript": Sequence({"chunk": Value("string"), "speaker": Value("string"), "text": Value("string")}),
                }
            )
        )

    def _split_generators(self, dl_manager):
        data_files = self.config.data_files

        return [
            SplitGenerator(
                name=Split.TRAIN,
                gen_kwargs={"filepath": data_files["train"][0]},
            ),
        ]

    def _generate_examples(self, filepath):
        df = pd.read_parquet(filepath)

        for idx, row in df.iterrows():
            record = row.to_dict()

            # Normalize vector field if it exists
            if "vector" in record:
                if isinstance(record["vector"], (list, tuple)):
                    record["vector"] = list(map(float, record["vector"]))
                else:
                    record["vector"] = []

            yield idx, record