beir.py

import json
import csv
import os
import datasets

logger = datasets.logging.get_logger(__name__)


_DESCRIPTION = "BEIR Benchmark"
_DATASETS = ["fiqa", "trec-covid", ""]

URL = ""
_URLs = {
    dataset: {
        "queries": URL + f"{dataset}/queries.jsonl", 
        "qrels": {
            "train": URL + f"{dataset}/qrels/train.tsv",
            "dev": URL + f"{dataset}/qrels/dev.tsv",
            "test": URL + f"{dataset}/qrels/test.tsv"
        }} for dataset in _DATASETS}


class BEIR(datasets.GeneratorBasedBuilder):
    """BEIR BenchmarkDataset."""

    BUILDER_CONFIGS = [
        datasets.BuilderConfig(
            name=dataset,
            description=f"This is the {dataset} dataset in BEIR Benchmark.",
        ) for dataset in _DATASETS
    ]
    

    def _info(self):
        return datasets.DatasetInfo(
            description=_DESCRIPTION,
             features=datasets.Features({
                "query": datasets.Value("string"), 
                "relevant": [{
                    "_id": datasets.Value("string"),
                    "score": datasets.Value("int32"),
                }],
             }),
            supervised_keys=None,
        )

    def _split_generators(self, dl_manager):
        """Returns SplitGenerators."""

        my_urls = _URLs[self.config.name]

        # All train, dev and test splits available for these datasets
        if self.config.name in ["msmarco", "nfcorpus", "hotpotqa", "fiqa", "fever"]:
            data_dir = dl_manager.download_and_extract(my_urls)
            return [
                datasets.SplitGenerator(
                    name=datasets.Split.TRAIN,
                    # These kwargs will be passed to _generate_examples
                    gen_kwargs={"query_path": data_dir["queries"],
                                "qrels_path": data_dir["qrels"]["train"]}
                ),
                datasets.SplitGenerator(
                    name="dev",
                    # These kwargs will be passed to _generate_examples
                    gen_kwargs={"query_path": data_dir["queries"],
                                "qrels_path": data_dir["qrels"]["dev"]}
                ),
                datasets.SplitGenerator(
                    name=datasets.Split.TEST,
                    # These kwargs will be passed to _generate_examples
                    gen_kwargs={"query_path": data_dir["queries"],
                                "qrels_path": data_dir["qrels"]["test"]}
                ),
            ]
        
        # Only train and test splits available for these datasets
        elif self.config.name in ["nq", "scifact"]:
            my_urls["qrels"].pop("dev", None)
            data_dir = dl_manager.download_and_extract(my_urls)
            
            return [
                datasets.SplitGenerator(
                    name=datasets.Split.TRAIN,
                    # These kwargs will be passed to _generate_examples
                    gen_kwargs={"query_path": data_dir["queries"],
                                "qrels_path": data_dir["qrels"]["train"]}
                ),
                datasets.SplitGenerator(
                    name=datasets.Split.TEST,
                    # These kwargs will be passed to _generate_examples
                    gen_kwargs={"query_path": data_dir["queries"],
                                "qrels_path": data_dir["qrels"]["test"]}
                ),
            ]
        
        # Only dev and test splits available for these datasets
        elif self.config.name in ["dbpedia", "quora"]:
            my_urls["qrels"].pop("train", None)
            data_dir = dl_manager.download_and_extract(my_urls)
            return [
                datasets.SplitGenerator(
                    name="dev",
                    # These kwargs will be passed to _generate_examples
                    gen_kwargs={"query_path": data_dir["queries"],
                                "qrels_path": data_dir["qrels"]["dev"]}
                ),
                datasets.SplitGenerator(
                    name=datasets.Split.TEST,
                    # These kwargs will be passed to _generate_examples
                    gen_kwargs={"query_path": data_dir["queries"],
                                "qrels_path": data_dir["qrels"]["test"]}
                ),
            ]
        
        # Only test split available for these datasets
        else:
            for split in ["train", "dev"]:
                my_urls["qrels"].pop(split, None)
            data_dir = dl_manager.download_and_extract(my_urls)
            return [
                datasets.SplitGenerator(
                    name=datasets.Split.TEST,
                    # These kwargs will be passed to _generate_examples
                    gen_kwargs={"query_path": data_dir["queries"],
                                "qrels_path": data_dir["qrels"]["test"]}
                ),
            ]


    def _generate_examples(self, query_path, qrels_path):
        """Yields examples."""
        
        queries, qrels = {}, {}

        with open(query_path, encoding="utf-8") as fIn:        
            text = fIn.readlines()

        for line in text:
            line = json.loads(line)
            queries[line.get("_id")] = line.get("text", "")
        
        reader = csv.reader(open(qrels_path, encoding="utf-8"), 
                    delimiter="\t", quoting=csv.QUOTE_MINIMAL)
        
        next(reader)

        for id, row in enumerate(reader):
            query_id, corpus_id, score = row[0], row[1], int(row[2])
            if query_id not in qrels:
                qrels[query_id] = {corpus_id: score}
            else:
                qrels[query_id][corpus_id] = score
        
        for i, query_id in enumerate(qrels):
            yield i, {
                "query": queries[query_id], 
                "relevant": [{"_id": doc_id, "score": score
                    } for doc_id, score in qrels[query_id].items()]
                    }