"""
Convert the Thresholding CSV files to Arrow format,
downloading the real files from HuggingFace (bypassing LFS pointers).
"""
from huggingface_hub import hf_hub_download
import pyarrow as pa
import pyarrow.csv as pcsv
from pathlib import Path

REPO = "AnnaWegmann/AV"
SPLITS = {
    "train": "thresholding/train.csv",
    "validation": "thresholding/validation.csv",
    "test": "thresholding/test.csv",
}

for split, csv_repo_path in SPLITS.items():
    print(f"\n--- {split} ---")

    # Download real CSV from HuggingFace
    local_csv = hf_hub_download(REPO, csv_repo_path, repo_type="dataset")
    print(f"  Downloaded: {local_csv}")

    # Read CSV into Arrow table (texts contain newlines)
    parse_opts = pcsv.ParseOptions(newlines_in_values=True)
    table = pcsv.read_csv(local_csv, parse_options=parse_opts)
    print(f"  Rows: {table.num_rows}, Cols: {table.column_names}")

    # Write as Arrow IPC streaming format (same as the working Contrastive_Learning files)
    out_dir = Path("thresholding") / split
    out_dir.mkdir(parents=True, exist_ok=True)
    out_path = out_dir / "data-00000-of-00001.arrow"

    with open(out_path, "wb") as f:
        writer = pa.ipc.new_stream(f, table.schema)
        writer.write_table(table)
        writer.close()

    print(f"  Wrote: {out_path} ({out_path.stat().st_size:,} bytes)")

print("\nDone! Now delete the old CSV files, update README.md, and push.")