generate_evaluation_datasets.py

import typer
from datasets import (Dataset, DatasetDict, get_dataset_config_names,
                      load_dataset)
from huggingface_hub import list_datasets
import pandas as pd

app = typer.Typer()


def convert(dataset_id: str):
    errors = []
    dataset_name = dataset_id.split("/")[-1]
    try:
        configs = get_dataset_config_names(dataset_id)
    except:
        typer.echo(f"❌ Failed to get configs for {dataset_id}")
        errors.append({"dataset_name": dataset_id, "error_type": "config"})
        return errors

    for config in configs:
        typer.echo(f"🛠️🛠️🛠️ Converting {dataset_id} with config {config} 🛠️🛠️🛠️")
        try:
            raw_datasets = load_dataset(dataset_id, name=config)
        except:
            typer.echo(f"❌ Failed to load {dataset_id} with config {config}")
            errors.append({"dataset_name": dataset_id, "config": config, "error_type": "load"})
            continue
        datasets_to_convert = DatasetDict()

        for split, dataset in raw_datasets.items():
            if split not in ["train", "validation"]:
                datasets_to_convert[split] = dataset

        for split, dataset in datasets_to_convert.items():
            columns_to_keep = ["gem_id", "target", "references"]
            remainder_cols = validate_columns(dataset)
            if len(remainder_cols) > 0:
                typer.echo(
                    f"❌ Skipping {dataset_name}/{config}/{split} due to missing columns: {', '.join(remainder_cols)}"
                )
                errors.append({"dataset_name": dataset_id, "config": config, "split": split, "error_type": "missing_columns", "missing_columns": remainder_cols})
            else:
                # Add `input` column if it exists
                if "input" in dataset.column_names:
                    columns_to_keep.append("input")
                # The test split doesn't have a parent ID
                # TODO(lewtun): check this logic!
                if split != "test" and "gem_parent_id" in dataset.column_names:
                    columns_to_keep.append("gem_parent_id")
                # The `datasets` JSON serializer is buggy - use `pandas` for now
                df = dataset.to_pandas()
                # Exclude dummy config names for comparison with GitHub source dataset
                if config in ["default", "xsum", "totto"]:
                    reference_name = f"{dataset_name}_{split}"
                else:
                    reference_name = f"{dataset_name}_{config}_{split}"
                df[columns_to_keep].to_json(f"{reference_name}.json", orient="records")

        typer.echo(f"✅ Successfully converted {dataset_id} with config {config}")

    return errors


def validate_columns(dataset: Dataset):
    ref_columns = ["gem_id", "target", "references"]
    columns = dataset.column_names
    return set(ref_columns) - set(columns)


@app.command()
def extract_evaluation_datasets():
    errors = []
    all_datasets = list_datasets()
    # Filter for GEM datasets
    gem_datasets = [dataset for dataset in all_datasets if dataset.id.startswith("GEM/")]
    # Filter for blocklist
    blocklist = [
        "indonlg", # Can't load
        "RiSAWOZ", # Can't load
        "CrossWOZ", # Can't load
        "references",  # This repo, so exclude!
    ]
    blocklist = ["GEM/" + dataset for dataset in blocklist]
    gem_datasets = [dataset for dataset in gem_datasets if dataset.id not in blocklist]
    for dataset in gem_datasets:
        errors.extend(convert(dataset.id))

    if len(errors):
        typer.echo("🙈 Found conversion errors!")
        errors_df = pd.DataFrame(errors)
        errors_df.to_csv("conversion_errors.csv", index=False)

    typer.echo(f"🥳 All datasets converted!")


if __name__ == "__main__":
    app()