services/analysis/clickbait_classifier.py

import os
from typing import Dict, Any

BASE_MODEL_NAME = "distilbert-base-uncased"

# Path to the LoRA adapter produced by training/train_clickbait_classifier.py
_MODEL_DIR = os.path.join(
    os.path.dirname(os.path.dirname(os.path.dirname(os.path.abspath(__file__)))),
    "models",
    "clickbait-lora",
)

_DEFAULT_RESULT: Dict[str, Any] = {
    "clickbait": False,
    "clickbait_score": 0.0,
    "model_status": "not_loaded",
    "explanation": "Finetuned clickbait model not found. Run "
                    "backend/training/train_clickbait_classifier.py and copy "
                    "the output into backend/models/clickbait-lora/ to enable "
                    "this feature.",
}

# Lazy-loaded globals - we only import torch/transformers and load
# weights the first time analyze_clickbait() is actually called, so a
# missing/broken model never slows down or crashes app startup.
_model = None
_tokenizer = None
_load_attempted = False
_load_error = None


def _try_load_model():
    global _model, _tokenizer, _load_attempted, _load_error

    if _load_attempted:
        return

    _load_attempted = True

    if not os.path.isdir(_MODEL_DIR):
        _load_error = f"Model directory not found at {_MODEL_DIR}"
        return

    try:
        import torch  # noqa: F401
        from transformers import AutoTokenizer, AutoModelForSequenceClassification
        from peft import PeftModel

        tokenizer = AutoTokenizer.from_pretrained(_MODEL_DIR)
        base_model = AutoModelForSequenceClassification.from_pretrained(
            BASE_MODEL_NAME,
            num_labels=2,
            id2label={0: "not_clickbait", 1: "clickbait"},
            label2id={"not_clickbait": 0, "clickbait": 1},
        )
        model = PeftModel.from_pretrained(base_model, _MODEL_DIR)
        model.eval()

        _model = model
        _tokenizer = tokenizer

    except Exception as e:  # noqa: BLE001 - intentionally broad, this must never crash the app
        _load_error = str(e)
        _model = None
        _tokenizer = None


def _get_headline(text: str) -> str:
    """Use the first sentence as a headline proxy if no explicit title exists."""
    if not text:
        return ""
    first_sentence = text.strip().split(".")[0]
    # Keep it short - headlines/titles are short by nature
    return first_sentence[:200]


def analyze_clickbait(text: str) -> Dict[str, Any]:
    """
    Returns a dict:
        {
            "clickbait": bool,
            "clickbait_score": float (0.0 - 1.0),
            "model_status": "loaded" | "not_loaded" | "error",
            "explanation": str
        }

    This function NEVER raises. Any failure path returns a safe default
    so callers (routes/analyze.py) don't need special-case error handling
    beyond what they already do for the other analyzers.
    """

    if not text or not text.strip():
        return {
            **_DEFAULT_RESULT,
            "model_status": "not_loaded" if _model is None else "loaded",
            "explanation": "Empty input text.",
        }

    _try_load_model()

    if _model is None or _tokenizer is None:
        result = dict(_DEFAULT_RESULT)
        if _load_error:
            result["explanation"] = (
                "Clickbait model unavailable, using neutral default. "
                f"(Reason: {_load_error})"
            )
        return result

    try:
        import torch

        headline = _get_headline(text)

        inputs = _tokenizer(
            headline,
            return_tensors="pt",
            truncation=True,
            max_length=64,
        )

        with torch.no_grad():
            outputs = _model(**inputs)
            probs = torch.softmax(outputs.logits, dim=-1)[0]

        clickbait_score = float(probs[1].item())

        return {
            "clickbait": clickbait_score >= 0.5,
            "clickbait_score": round(clickbait_score, 4),
            "model_status": "loaded",
            "explanation": (
                f"Finetuned model classified this as "
                f"{'likely' if clickbait_score >= 0.5 else 'unlikely'} clickbait "
                f"(confidence {round(clickbait_score, 2)})."
            ),
        }

    except Exception as e:  # noqa: BLE001 - never let inference errors reach the caller
        return {
            **_DEFAULT_RESULT,
            "model_status": "error",
            "explanation": f"Clickbait model inference failed, using neutral default. ({e})",
        }


if __name__ == "__main__":
    samples = [
        "You Won't Believe What Happened Next!",
        "Central Bank Raises Interest Rates by 0.25 Percent",
    ]
    for s in samples:
        print(s, "->", analyze_clickbait(s))