build.py

import argparse
import polars as pl
import os

MAPPING = {"E": 3, "S": 2, "C": 1, "I": 0}
COLUMNS = [
    "product_description",
    "product_bullet_point",
    "product_brand",
    "product_color",
]

if __name__ == "__main__":
    parser = argparse.ArgumentParser()
    parser.add_argument(
        "--source", type=str, required=True, help="path to parquet files"
    )
    parser.add_argument("--small", action="store_true", help="select small subset")
    parser.add_argument("--split", type=str, choices=["train", "test"], required=True)
    parser.add_argument(
        "--locale",
        type=str,
        default="us",
        choices=["us", "es", "jp"],
        help="language to select",
    )

    args = parser.parse_args()

    products = pl.read_parquet(
        os.path.join(args.source, "shopping_queries_dataset_products.parquet")
    )

    examples = pl.read_parquet(
        os.path.join(args.source, "shopping_queries_dataset_examples.parquet")
    )
    merged = examples.join(products, on=pl.col("product_id"))

    merged = merged.select(pl.all().exclude("^__index_level_.*$"))
    merged = merged.with_columns(
        label=pl.col("esci_label").replace(MAPPING).cast(pl.Int32),
        id=pl.col("example_id").cast(pl.String),
        query_id=pl.col("query_id").cast(pl.String),
        product_id=pl.col("product_id").cast(pl.String),
        title=pl.col("product_title"),
        text=pl.concat_str(
            [pl.lit(f"{col}: ") + pl.col(col).fill_null("") for col in COLUMNS],
            separator="\n",
        ),
    )

    print(f"loaded {len(merged)} source rows")
    merged = merged.filter(pl.col("split") == args.split)
    print(f"split filtering done: {len(merged)} rows")
    merged = merged.filter(pl.col("small_version") == args.small)
    print(f"size filtering done: {len(merged)} rows")
    merged = merged.filter(pl.col("product_locale") == args.locale)
    print(f"locale filtering done: {len(merged)} rows")
    merged.write_ndjson(f"{args.split}.jsonl")