"""Fälle und Präferenzpaare: modellneutral speichern, deterministisch splitten, exportieren."""

from __future__ import annotations

import hashlib
import json
from collections.abc import Iterator
from typing import Any

from .store import Store

EVAL_SHARE_PERCENT = 20


def split_for(text: str) -> str:
    """Fester Split pro Eingabe: gleiche Eingabe landet immer im selben Split."""
    bucket = int(hashlib.sha256(text.encode("utf-8")).hexdigest(), 16) % 100
    return "eval" if bucket < EVAL_SHARE_PERCENT else "train"


def export_sft(store: Store, *, cluster: str, split: str = "train") -> Iterator[str]:
    """JSONL für Supervised Finetuning: nur unabhängig verifizierte Fälle."""
    for case in reversed(store.list_cases(cluster=cluster, split=split, passed=True)):
        messages = list(case["messages"] or [])
        messages.append({"role": "assistant", "content": case["output"]})
        meta = {k: case[k] for k in ("id", "skill", "skill_version", "stage", "model", "origin")}
        yield json.dumps({"messages": messages, "meta": meta}, ensure_ascii=False)


def export_preferences(store: Store, *, cluster: str, split: str = "train") -> Iterator[str]:
    for pair in store.list_preference_pairs(cluster=cluster):
        if pair["split"] != split:
            continue
        yield json.dumps(
            {"prompt": pair["prompt"], "chosen": pair["chosen"], "rejected": pair["rejected"]},
            ensure_ascii=False,
        )


def pass_rate(rows: list[dict[str, Any]]) -> float | None:
    decided = [r for r in rows if r.get("passed") is not None]
    if not decided:
        return None
    return sum(1 for r in decided if r["passed"]) / len(decided)