"""Rebuild the verified MedAraBench test set from QIMMA's public copy.

Writes one file per config, <output_dir>/<config>/data.parquet, with the
source columns unchanged.

Usage:
    python apply_filter.py [output_dir]
"""
import hashlib
import json
import os
import sys

import pandas as pd
from huggingface_hub import hf_hub_download

SOURCE = "qimma/MCQ_MedAraBench"
SOURCE_REVISION = "c62cc54f8d9d9e9e4be5aca82982b58e2da3cf0e"
THIS_REPO = "IFM/MedAraBench_verified"
COLUMNS = ["prompt", "question", "choices", "index"]
EXPECTED_KEPT = {
    "Anatomy": 830, "Histology": 188, "Histology and Anatomy": 39,
    "Histology and Systemic Anatomy": 8, "Anesthesia": 70, "Biochemistry": 303,
    "Cell Biology": 197, "Genetics": 176, "Physics": 521, "Chemistry": 71,
    "Embryology": 5, "Emergency Medicine": 21, "Gastroenterology": 29,
    "Neurology": 43, "Cardiology": 17, "Thorasic Internal": 17, "Oncology": 12,
    "Chest Diseases": 6, "Gastroenterology & Hepatology": 38, "Microbiology": 115,
    "Ophthalmology": 234, "Pathology": 43, "Pharmacology": 62, "Physiology": 573,
    "Statistics": 310, "Surgery": 103,
}


def fingerprint(row):
    values = [row["prompt"], row["question"],
              json.dumps(list(row["choices"]), ensure_ascii=False), str(row["index"])]
    return hashlib.sha256("\x1f".join(values).encode()).hexdigest()[:16]


ids = pd.read_csv(hf_hub_download(THIS_REPO, "ids.csv", repo_type="dataset"))
out = sys.argv[1] if len(sys.argv) > 1 else "medarabench_verified"
total = 0
for config, expected in EXPECTED_KEPT.items():
    cfg = ids[ids["config"].eq(config)].sort_values("row").reset_index(drop=True)
    source = pd.read_parquet(hf_hub_download(SOURCE, f"{config}/data.parquet",
                                             repo_type="dataset", revision=SOURCE_REVISION))
    assert len(source) == len(cfg), f"{config}: expected {len(cfg)} source rows, got {len(source)}"
    mismatch = [i for i, row in source.iterrows() if fingerprint(row) != cfg.at[i, "fingerprint"]]
    assert not mismatch, f"{config}: source rows differ from the audited version: {mismatch[:10]}"

    verified = source[cfg["status"].eq("kept").to_numpy()].reset_index(drop=True)
    assert len(verified) == expected, f"{config}: expected {expected} rows, got {len(verified)}"
    os.makedirs(os.path.join(out, config), exist_ok=True)
    verified.to_parquet(os.path.join(out, config, "data.parquet"))
    total += len(verified)

assert total == 4031, f"expected 4031 rows, got {total}"
print(f"Wrote {total} rows in {len(EXPECTED_KEPT)} configs to {out}/")