birco.py

import os
import datasets
from datasets import DownloadManager, DatasetInfo, Features, Value, Split, SplitGenerator

_CITATION = """"@misc{birco,\n   title={{BIRCO: A Benchmark of Information Retrieval Tasks with Complex Objectives}},\n   author={{Xiaoyue Wang et al.}},\n   year={2024},\n   url={https://arxiv.org/abs/2402.14151},\n}""""
_DESCRIPTION = """"BIRCO benchmark containing corpus, queries, and relevance judgments""""
_HOMEPAGE = "https://github.com/BIRCO-benchmark/BIRCO"
_LICENSE = "CC-BY-4.0"

class BIRCO(datasets.GeneratorBasedBuilder):
    BUILDER_CONFIGS = [
        datasets.BuilderConfig(
            name="corpus",
            version=datasets.Version("1.0.0"),
            description="Document corpus",
        ),
        datasets.BuilderConfig(
            name="queries",
            version=datasets.Version("1.0.0"),
            description="Search queries",
        ),
        datasets.BuilderConfig(
            name="default",
            version=datasets.Version("1.0.0"),
            description="Relevance judgments",
        ),
    ]

    def _info(self):
        if self.config.name == "corpus":
            features = Features({
                "_id": Value("string"),
                "text": Value("string"),
                "title": Value("string")
            })
        elif self.config.name == "queries":
            features = Features({
                "_id": Value("string"),
                "text": Value("string")
            })
        elif self.config.name == "default":
            features = Features({
                "query-id": Value("string"),
                "corpus-id": Value("string"),
                "score": Value("float64")
            })

        return DatasetInfo(
            description=_DESCRIPTION,
            features=features,
            citation=_CITATION,
            homepage=_HOMEPAGE,
            license=_LICENSE
        )

    def _split_generators(self, dl_manager):
        return [
            SplitGenerator(
                name=Split.TEST,
                gen_kwargs={
                    "files": dl_manager.download_and_extract({
                        "data": f"data/{self.config.name}/test-*.parquet"
                    }),
                    "split": "test"
                }
            )
        ]

    def _generate_examples(self, files, split):
        dataset = datasets.load_dataset("parquet", data_files=files["data"], split=split)
        for idx, example in enumerate(dataset):
            yield idx, example