testbed/embeddings-benchmark__mteb/scripts/data/wikipedia_retrieval/create_data.py

from time import sleep

from datasets import DatasetDict, Features, Value, load_dataset
from huggingface_hub import HfApi, hf_hub_download

languages = [
    "de",
    "bn",
    "it",
    "pt",
    "nl",
    "cs",
    "ro",
    "bg",
    "sr",
    "fi",
    "fa",
    "hi",
    "da",
    "en",
    "no",
    "sv",
]


def apply_query_id(example, queries_dict):
    title = example["title"]
    corpus_id = example["_id"]
    score = example["score"]
    query_id = queries_dict[title]
    return {"query-id": query_id, "corpus-id": corpus_id, "score": score}


for lang in languages:
    ds_queries = load_dataset(
        f"rasdani/cohere-wikipedia-2023-11-{lang}-queries", split="train"
    )
    ds_corpus = load_dataset(
        f"rasdani/cohere-wikipedia-2023-11-{lang}-1.5k-articles", split="train"
    )
    sleep(5)  # HF hub rate limit

    queries = ds_queries.map(
        lambda x: {"_id": "q" + x["_id"], "text": x["query"]},
        remove_columns=ds_queries.column_names,
    )
    queries_dict = {row["title"]: "q" + row["_id"] for row in ds_queries}
    corpus = ds_corpus.map(
        lambda x: {"_id": x["_id"], "title": x["title"], "text": x["text"]},
        remove_columns=ds_corpus.column_names,
    )
    qrels = ds_corpus.map(
        lambda x: apply_query_id(x, queries_dict=queries_dict),
        remove_columns=ds_corpus.column_names,
    )

    corpus_features = Features(
        {"_id": Value("string"), "title": Value("string"), "text": Value("string")}
    )
    queries_features = Features({"_id": Value("string"), "text": Value("string")})
    qrels_features = Features(
        {
            "query-id": Value("string"),
            "corpus-id": Value("string"),
            "score": Value("float32"),
        }
    )

    corpus = corpus.cast(corpus_features)
    queries = queries.cast(queries_features)
    qrels = qrels.cast(qrels_features)

    ds_dict = DatasetDict({"corpus": corpus, "queries": queries, "qrels": qrels})

    # repo_id = f"ellamind/wikipedia-2023-11-retrieval-{lang}"
    repo_id = "ellamind/wikipedia-2023-11-retrieval-multilingual"
    # corpus.push_to_hub(repo_id, config_name="corpus", split="test")
    # queries.push_to_hub(repo_id, config_name="queries", split="test")
    # qrels.push_to_hub(repo_id, config_name="default", split="test")
    corpus.push_to_hub(repo_id + "-corpus", config_name=lang, split="test")
    queries.push_to_hub(repo_id + "-queries", config_name=lang, split="test")
    qrels.push_to_hub(repo_id + "-qrels", config_name=lang, split="test")

    # Download the README from the repository
    sleep(5)
    readme_path = hf_hub_download(
        repo_id=repo_id, filename="README.md", repo_type="dataset"
    )

    with open(readme_path, "r") as f:
        readme_content = f.read()

    readme = """
This dataset is derived from Cohere's wikipedia-2023-11 dataset, which is in turn derived from `wikimedia/wikipedia`.
The dataset is licensed under the Creative Commons CC BY-SA 3.0 license.
    """
    # Prepend the license key to the YAML header and append the custom README
    if "- license: " not in readme_content and readme not in readme_content:
        license = "cc-by-sa-3.0"

        updated_readme = readme_content.replace(
            "---\ndataset_info:", "---\nlicense: {license}\ndataset_info:"
        ).format(license=license)
        updated_readme += readme

        api = HfApi()
        readme_bytes = updated_readme.encode("utf-8")
        api.upload_file(
            path_or_fileobj=readme_bytes,
            path_in_repo="README.md",
            repo_id=repo_id,
            repo_type="dataset",
        )