load.py

import json
import os
import datasets

# Define the dataset
class M3Retrieve(datasets.GeneratorBasedBuilder):
    VERSION = datasets.Version("1.0.0")

    # Define available subfolders (datasets)
    SUBFOLDERS = [
        "Anatomy and Physiology",
        "Cardiology",
        "Dermatology",
        "Endocrinology_and_Diabetes",
        "Gastroenterology",
        "Hematology",
        "Microbiology_and_Cell_Biology",
        "Miscellaneous",
        "Neurology_and_Neuroscience",
        "Ophthalmology_and_Sensory_Systems",
        "Orthopedics_and_Musculoskeletal",
        "Pharmacology",
        "Psychiatry_and_Mental_Health",
        "Pubmed",
        "Radiology_and_Imaging",
        "Reproductive_System",
        "Respiratory_and_Pulmonology",
        "Surgical_Specialties",
    ]

    # Define Builder Configs for each subfolder
    BUILDER_CONFIGS = [
        datasets.BuilderConfig(
            name=subfolder,
            version=datasets.Version("1.0.0"),
            description=f"Dataset for {subfolder.replace('_', ' ')}"
        )
        for subfolder in SUBFOLDERS
    ]

    # def _info(self):
    #     return datasets.DatasetInfo(
    #         description="M3Retrieve: Benchmarking Multimodal Retrieval for Medicine",
    #         features={
    #             "queries": {
    #                 "_id": datasets.Value("string"),
    #                 "caption": datasets.Value("string"),
    #                 "image_path": datasets.Value("string"),
    #             },
    #             "corpus": {
    #                 "_id": datasets.Value("string"),
    #                 "text": datasets.Value("string"),
    #             },
    #             "qrels": {
    #                 "query-id": datasets.Value("string"),
    #                 "corpus-id": datasets.Value("string"),
    #                 "score": datasets.Value("float32"),
    #             },
    #         },
    #         supervised_keys=None,
    #     )

    def _info(self):
        return datasets.DatasetInfo(
            description="M3Retrieve: Benchmarking Multimodal Retrieval for Medicine",
            features=datasets.Features(
                {
                    "_id": datasets.Value("string"),
                    "caption": datasets.Value("string"),
                    "image_path": datasets.Value("string"),
                    "text": datasets.Value("string"),
                    "query-id": datasets.Value("string"),
                    "corpus-id": datasets.Value("string"),
                    "score": datasets.Value("float32"),
                }
            ),
            supervised_keys=None,
        )


    def _split_generators(self, dl_manager):
        """Returns SplitGenerators for the selected subfolder"""
        data_dir = os.path.join(dl_manager.download_and_extract(self.config.data_dir), self.config.name)

        return [
            datasets.SplitGenerator(
                name="queries",
                gen_kwargs={"filepath": os.path.join(data_dir, "queries.jsonl"), "key": "queries"},
            ),
            datasets.SplitGenerator(
                name="corpus",
                gen_kwargs={"filepath": os.path.join(data_dir, "corpus.jsonl"), "key": "corpus"},
            ),
            datasets.SplitGenerator(
                name="qrels",
                gen_kwargs={"filepath": os.path.join(data_dir, "qrels/test.tsv"), "key": "qrels"},
            ),
        ]

    def _generate_examples(self, filepath, key):
        """Yields examples as (key, example) tuples."""
        if key in ["queries", "corpus"]:
            with open(filepath, "r", encoding="utf-8") as f:
                for i, line in enumerate(f):
                    data = json.loads(line)
                    yield i, data
        elif key == "qrels":
            with open(filepath, "r", encoding="utf-8") as f:
                for i, line in enumerate(f):
                    query_id, corpus_id, score = line.strip().split("\t")
                    yield i, {"query-id": query_id, "corpus-id": corpus_id, "score": float(score)}