MixBench25.py

import os
import json
import datasets
import csv 

_DESCRIPTION = """\
MixBench is a benchmark for evaluating mixed-modality retrieval. It contains queries and corpora from four datasets: MSCOCO, Google_WIT, VisualNews, and OVEN. \
Each subset provides: query, corpus, mixed_corpus, and qrel splits.
"""


_HOMEPAGE = "https://huggingface.co/datasets/mixed-modality-search/MixBench25"


_SUBSETS = ["MSCOCO", "Google_WIT", "VisualNews", "OVEN"]

class MixBenchConfig(datasets.BuilderConfig):
    def __init__(self, name, **kwargs):
        if name not in _SUBSETS:
            raise ValueError(f"Unknown subset: {name}. Choose from {_SUBSETS}")
        super().__init__(name=name, version=datasets.Version("1.0.0"), **kwargs)


class MixBench(datasets.GeneratorBasedBuilder):
    BUILDER_CONFIGS = [MixBenchConfig(name=subset) for subset in _SUBSETS]

    def _info(self):
        features = datasets.Features({
            "query_id": datasets.Value("string"),
            "corpus_id": datasets.Value("string"),
            "text": datasets.Value("string"),
            "image": datasets.Value("string"),
            "score": datasets.Value("int32"),
        })
        return datasets.DatasetInfo(
            description=_DESCRIPTION,
            features=features,
            homepage=_HOMEPAGE,
        )

    def _split_generators(self, dl_manager):
        # download entire repo root and go to current subset folder
        # data_dir = dl_manager.download_and_extract(".")
        # subset_dir = os.path.join(data_dir, self.config.name)
        # repo_root = dl_manager.manual_dir or os.path.join(os.path.dirname(__file__), self.config.name)
        # subset_dir = os.path.join(repo_root, self.config.name)
        subset_dir = os.path.join(dl_manager.manual_dir or dl_manager._base_path, self.config.name)
        return [
            datasets.SplitGenerator(
                name="query",
                gen_kwargs={"path": os.path.join(subset_dir, "queries.jsonl"), "split": "query"},
            ),
            datasets.SplitGenerator(
                name="corpus",
                gen_kwargs={"path": os.path.join(subset_dir, "corpus.jsonl"), "split": "corpus"},
            ),
            datasets.SplitGenerator(
                name="mixed_corpus",
                gen_kwargs={"path": os.path.join(subset_dir, "mixed_corpus.jsonl"), "split": "mixed_corpus"},
            ),
            datasets.SplitGenerator(
                name="qrel",
                gen_kwargs={"path": os.path.join(subset_dir, "qrels", "qrels.tsv"), "split": "qrel"},
            ),
        ]

    def _generate_examples(self, path, split):
        if split == "qrel":
            # with open(path, encoding="utf-8") as f:
            #     for idx, line in enumerate(f):
            #         qid, did, score = line.strip().split()
            #         yield idx, {
            #             "query_id": qid,
            #             "corpus_id": did,
            #             "text": "",
            #             "image": "",
            #             "score": int(score),
            #         }
            with open(path, encoding="utf-8") as f:
                reader = csv.DictReader(f, delimiter="\t")  # 使用 DictReader 读取有表头的 tsv
                for idx, row in enumerate(reader):
                    yield idx, {
                        "query_id": row["query_id"],
                        "corpus_id": row["corpus_id"],
                        "score": int(row["score"]),
                    }
        else:
            with open(path, encoding="utf-8") as f:
                for idx, line in enumerate(f):
                    row = json.loads(line)
                    yield idx, {
                        "query_id": row.get("query_id", ""),
                        "corpus_id": row.get("corpus_id", ""),
                        "text": row.get("text", ""),
                        "image": row.get("image", ""),
                        "score": 0,
                    }