PITTI/wine-reviews

Dataset adapted from spawn99/wine-reviews to train classifiers on grape variety.
Columns were consolidated to match the format described in this project

import polars as pl
from datasets import load_dataset

ds_dict = load_dataset("spawn99/wine-reviews")

processed_splits = {}

for split_name, ds in ds_dict.items():
  print(f"Processing {split_name} split...")
  
  # Convert to Polars (Zero-copy via Arrow)
  df = pl.from_arrow(ds.data.table)

  # Apply transformation logic
  df_final = df.select([
    pl.format(
      "Based on this wine review, guess the grape variety:\n"
      "This wine is produced by {} in the {} region of {}.\n"
      "It was grown in {}. It is described as: \"{}\".\n"
      "The wine has been reviewed by {} and received {} points.\n"
      "The price is {}.",
      pl.col("winery").fill_null("a winery"),
      # Region logic: region_1 or province or region_2 or "Unknown region"
      pl.coalesce(["region_1", "province", "region_2"]).fill_null("Unknown region"),
      pl.col("country").fill_null("Unknown country"),
      pl.col("designation").fill_null("an unspecified appellation"),
      pl.col("description").fill_null("No description provided."),
      pl.col("taster_name").fill_null("a reviewer"),
      pl.col("points").cast(pl.String).fill_null("unrated"),
      # Price logic: cast to int to remove .0 then to string
      pl.col("price").cast(pl.Int64).cast(pl.String).fill_null("unknown")
    ).alias("text"),
    
    pl.col("variety").alias("label")
  ]).filter(pl.col("label").is_not_null())

  processed_splits[split_name] = df_final
  
  # Save locally or inspect
  print(f"Split {split_name} finished. Rows: {len(df_final)}")
  df_final.write_parquet(f"processed/{split_name}.parquet")

Original Dataset Details

• License: CC BY-NC-SA 4.0
• Attribution: Zackthoutt
• Source: Wine Reviews Dataset on Kaggle