"""Pure UAReviews intent adapter. No downloading, translation or training.

The Hub exposes one physical train file containing THREE logical splits. Always
use source_split returned here, never the Hub container split. Keep test and
challenge exclusively for evaluation, including when deduplicating against train.
The caller must group duplicate texts across splits and give evaluation priority.
Human final labels are categorical judgments, not measured probabilities. Only
20% of the source received the paper's three-annotator gold validation; the file
does not identify that subset, so no row is claimed to be independently verified.
"""

import unicodedata


SOURCE = {
    "repo": "KSE-RESEARCH-Group/UAReviews",
    "revision": "6f2ad474981e453fdf483dee1a6db060888bb212",
    "license": "cc-by-4.0",
    "file": "benchmark_v1_splits.jsonl",
    "card": "https://huggingface.co/datasets/KSE-RESEARCH-Group/UAReviews/blob/"
    "6f2ad474981e453fdf483dee1a6db060888bb212/README.md",
    "paper": "https://aclanthology.org/2026.unlp-1.2/",
}
# Exact source taxonomy. Candidate wording stays faithful to the published card;
# no invented detailed decision rules or translated Ukrainian input are added.
CANDIDATES = (
    "Gratitude / Positive Feedback",
    "Complaint / Dissatisfaction",
    "Question / Request for Help",
    "Suggestion / Idea",
    "Neutral Comment",
)
KEYS = ("gratitude", "complaint", "question", "suggestion", "neutral")
SPLIT_COUNTS = {"train": 8106, "test": 1737, "challenge": 1737}
INSTRUCTIONS = (
    "Classify the primary communicative intent of this Ukrainian user review or "
    "feedback comment using the five provided categories. Select one category."
)


def adapt(record, expected_split=None):
    """Return zero/one Choice task; fail closed on absent/unknown logical splits.

    expected_split optionally asserts a caller's logical partition; it must not
    be passed as 'train' simply because HF loaded a physical train container.
    Unknown labels are errors rather than silently reinterpreted classifications.
    Source ratings, emotions, IDs, labels and provenance stay audit-only.
    """
    if not isinstance(record, dict):
        return []
    split = record.get("split")
    if split not in SPLIT_COUNTS:
        raise ValueError("UAReviews requires a train/test/challenge row-level split")
    if expected_split is not None and expected_split != split:
        raise ValueError("UAReviews row-level split differs from requested partition")
    label = record.get("final_category")
    if label not in CANDIDATES:
        raise ValueError("Unknown UAReviews intent label: %r" % label)
    content = record.get("content")
    if not isinstance(content, str) or not content.strip():
        return []
    state = content.replace("\r\n", "\n").replace("\r", "\n")
    group_key = " ".join(unicodedata.normalize("NFC", state).casefold().split())
    return [{
        "state": state,
        "instructions": INSTRUCTIONS,
        "candidates": list(CANDIDATES),
        "keys": list(KEYS),
        "target": [float(candidate == label) for candidate in CANDIDATES],
        "kind": "choice",
        "group_key": group_key,
        "source_split": split,
        "split": split,
        "metadata": {
            "source": SOURCE["repo"],
            "source_revision": SOURCE["revision"],
            "source_license": SOURCE["license"],
            "source_card": SOURCE["card"],
            "source_paper": SOURCE["paper"],
            "source_file": SOURCE["file"],
            "source_row_id": record.get("id"),
            "source_split": split,
            "evaluation_only": split != "train",
            "source_origin": record.get("source"),
            "language": "uk",
            "label_method": "published_final_human_intent_annotation",
            "target_semantics": "one_hot_categorical_annotation_not_calibrated_probability",
            "independent_gold_subset_membership": "not_identified_in_source",
        },
    }]