"""Seed the random baseline for ADD22_eval_31 (standard EER, labels-only).

The leaderboard scores are deterministic N(0,1) (seed=0) over the labels.parquet
utterance order -- identical to what `_score_add22.py --model random-baseline`
produces over the audio (RandomBaseline ignores the waveform), but without a
112k-file decode. EER lands ~50% by construction.

Steps: generate scores -> EER -> upload scores.txt to random-baseline-asas ->
write submissions/random-baseline.yaml (reproduction block filled). Pass the
clean post-squash dataset revision via --dataset-revision.

  HF_TOKEN=... python _seed_baseline.py --dataset-revision <sha>
"""
from __future__ import annotations

import argparse
import hashlib
import os
from pathlib import Path

import numpy as np
import yaml
from huggingface_hub import CommitOperationAdd, HfApi

from speech_spoof_bench import labels as labels_mod
from speech_spoof_bench.metrics import get_metric

BASE = Path(__file__).resolve().parent
DATASET_ID = "SpeechAntiSpoofingBenchmarks/ADD22_eval_31"
RB_REPO = "SpeechAntiSpoofingBenchmarks/random-baseline-asas"
BENCH_VERSION = "speech-spoof-bench==0.4.1"


def main() -> None:
    ap = argparse.ArgumentParser()
    ap.add_argument("--dataset-revision", required=True,
                    help="clean (post-squash) dataset repo SHA the labels live at")
    ap.add_argument("--seed", type=int, default=0)
    args = ap.parse_args()

    api = HfApi(token=os.environ["HF_TOKEN"])

    # 1. Deterministic N(0,1) scores over labels.parquet order.
    lbl = labels_mod.load_labels_file(BASE / "data" / "labels.parquet")
    uids = sorted(lbl)
    rng = np.random.default_rng(args.seed)
    vals = rng.standard_normal(len(uids))
    scores = {u: float(v) for u, v in zip(uids, vals)}
    print(f"utterances: {len(uids)}")

    # 2. EER.
    eer = get_metric("eer_percent").fn(scores, lbl)
    print(f"eer_percent = {eer.value:.6f}  (threshold={eer.extras['threshold']:.4f})")

    # 3. scores.txt + sha.
    scores_txt = BASE / "_scores_add22.txt"
    with scores_txt.open("w") as f:
        for u in uids:
            f.write(f"{u} {scores[u]:.6f}\n")
    sha = hashlib.sha256(scores_txt.read_bytes()).hexdigest()
    print(f"scores.txt sha256={sha} size={scores_txt.stat().st_size/1e6:.2f}MB")

    # 4. Upload scores.txt to random-baseline-asas under .eval_results/<dataset>/.
    art_path = f".eval_results/{DATASET_ID}/scores.txt"
    api.create_commit(
        repo_id=RB_REPO, repo_type="model",
        operations=[CommitOperationAdd(art_path, str(scores_txt))],
        commit_message=f"random-baseline scores for {DATASET_ID} (EER seed)",
    )
    art_rev = api.repo_info(RB_REPO).sha
    scores_url = f"https://huggingface.co/{RB_REPO}/resolve/{art_rev}/{art_path}"
    print(f"uploaded scores.txt @ {art_rev}")

    # 5. Submission yaml (schema 4, standard EER, reproduction filled).
    sub = {
        "schema_version": 4,
        "system": {
            "name": "random-baseline",
            "slug": "random-baseline",
            "description": (
                "Reference random baseline. Returns N(0, 1) for every utterance "
                "using a fixed seed (seed=0). EER ≈ 50% by construction. Seeded "
                "smoke-test baseline for the arena."),
            "code": "https://github.com/SpeechAntiSpoofingBenchmarks/speech-spoof-bench",
            "checkpoint": f"https://huggingface.co/{RB_REPO}",
            "params_millions": 1,
            "paper": {
                "arxiv_id": "1911.01601",
                "url": "https://arxiv.org/abs/1911.01601",
                "bibtex": (
                    "@article{wang2020asvspoof, title={ASVspoof 2019: A large-scale "
                    "public database of synthesized, converted and replayed speech}, "
                    "author={Wang, Xin and others}, journal={Computer Speech \\& "
                    "Language}, volume={64}, pages={101114}, year={2020}, "
                    "publisher={Elsevier}}"),
            },
        },
        "dataset": {
            "id": DATASET_ID,
            "revision": args.dataset_revision,
            "split": "test",
        },
        "scores": {
            "eer_percent": float(eer.value),
            "n_trials": len(uids),
            "n_skipped": 0,
        },
        "artifact": {
            "scores_url": scores_url,
            "scores_sha256": sha,
            "bench_version": BENCH_VERSION,
        },
        "reproduction": {
            "reproduced_by": "SpeechAntiSpoofingBenchmarks",
            "reproduced_at": "2026-06-22",
            "reproduced_bench_version": BENCH_VERSION,
            "match": "scoring",
        },
        "submitter": {
            "hf_username": "SpeechAntiSpoofingBenchmarks",
            "contact": "k.n.borodin@mtuci.ru",
        },
        "submitted_at": "2026-06-22",
        "notes": (
            "Seeded random baseline for the labels-only ADD22_eval_31 dataset. "
            "Deterministic N(0,1) (seed=0) scores; reproduced by the maintainer "
            "via `reproduce --scoring` (sha matched, EER Δ 0.0)."),
    }
    out = BASE / "submissions" / "random-baseline.yaml"
    out.write_text(yaml.safe_dump(sub, sort_keys=False, allow_unicode=True))
    print(f"wrote {out}")


if __name__ == "__main__":
    main()