cde_benchmark/formatters/data_formatter.py

from typing import List, Tuple
from datasets import DatasetDict, Dataset, load_dataset


class BaseDataFormatter:
    def get_nested(self) -> Tuple[List[List[str]], List[List[str]]]:
        raise NotImplementedError

    def get_flattened(self) -> Tuple[List[str], List[str]]:
        raise NotImplementedError

    def get_queries(self) -> Tuple[List[str], List[str]]:
        raise NotImplementedError


class DataFormatter(BaseDataFormatter):
    def __init__(self, dataset_path, split, query_key="queries", doc_key="documents"):
        self.doc_dataset = None
        self.queries_dataset = None
        self._load_from_path(dataset_path, split, query_key, doc_key)
        self.doc_dataset = self.doc_dataset.map(self.parse_id)

    def _load_from_path(self, path, split, query_key, doc_key):
        self.doc_dataset = load_dataset(path, doc_key, split=split)
        self.queries_dataset = load_dataset(path, query_key, split=split)
        # mapping dataset is used to map queries to relevant documents

    @staticmethod
    def parse_id(sample):
        doc_id, internal_id = sample["chunk_id"].split("_")
        return {"doc_id": doc_id, "internal_id": int(internal_id)}

    def get_nested(self) -> Tuple[List[List[str]], List[List[str]]]:
        # TODO: verify it's sorted
        return list(
            self.doc_dataset.to_pandas().groupby("doc_id")["chunk"].apply(list)
        ), list(self.doc_dataset.to_pandas().groupby("doc_id")["chunk_id"].apply(list))

    def get_flattened(self) -> Tuple[List[str], List[str]]:
        # flatten data
        return self.doc_dataset["chunk"], self.doc_dataset["chunk_id"]

    def get_queries(self) -> Tuple[List[str], List[str]]:
        return self.queries_dataset["query"], self.queries_dataset["chunk_id"]


class BEIRDataFormatter(BaseDataFormatter):
    def __init__(
        self,
        dataset_path,
        split,
        query_key="queries",
        doc_key="corpus",
        concat_num_docs=2,
    ):
        self.doc_dataset = None
        self.queries_dataset = None
        self.mapping = None
        self._load_from_path(dataset_path, split, query_key, doc_key)
        self.concat_num_docs = concat_num_docs

    def _load_from_path(self, path, split, query_key, doc_key):
        self.doc_dataset = load_dataset(path, doc_key, split=split)
        self.queries_dataset = load_dataset(path, query_key, split=split)
        mapping_dataset = load_dataset(path, "qrels", split=split)
        self.mapping = {
            query["query-id"]: query["corpus-id"] for query in mapping_dataset
        }
        # mapping dataset is used to map queries to relevant documents

    def get_nested(self) -> Tuple[List[List[str]], List[List[str]]]:

        self.doc_dataset = self.doc_dataset.shuffle(seed=42)
        idx = []
        for i in range(0, len(self.doc_dataset)):
            idx.extend([i] * self.concat_num_docs)
        idx = idx[: len(self.doc_dataset)]
        self.doc_dataset = self.doc_dataset.add_column("doc_id", idx)

        return list(
            self.doc_dataset.to_pandas().groupby("doc_id")["text"].apply(list)
        ), list(self.doc_dataset.to_pandas().groupby("doc_id")["_id"].apply(list))

    def get_flattened(self) -> Tuple[List[str], List[str]]:
        # flatten data
        return self.doc_dataset["text"], self.doc_dataset["_id"]

    def get_queries(self) -> Tuple[List[str], List[str]]:
        gold_docs = []
        for query in self.queries_dataset:
            gold_docs.append(self.mapping[query["_id"]])
        return self.queries_dataset["text"], gold_docs