README.md

---
library_name: sklearn
tags:
- text-classification
- sentence-transformers
- spoiler-detection
- sklearn
- movie-reviews
license: mit
---

# Multi-Source Spoiler Detector

This repository contains the trained classifier for a three-level movie-review spoiler detector.

## Task

The model predicts one of three labels:

- `Safe`: no meaningful spoiler detected
- `Mild`: broad setup, tone, or non-critical plot information
- `Major`: key twist, death, identity, ending, solution, or final outcome revealed

## Model

- Classifier: SVM with RBF kernel (`sklearn.svm.SVC`)
- Embeddings: `sentence-transformers/all-mpnet-base-v2`
- Input: English movie-review text
- Output: `Major`, `Mild`, or `Safe`

The serialized model is stored in `best_model.joblib`. It contains both the trained classifier and metadata with the embedding model name and label classes.

## Test Results

| Model | Accuracy | Macro F1 | Weighted F1 |
|---|---:|---:|---:|
| SVM RBF | 0.5753 | 0.5723 | 0.5752 |
| Logistic Regression | 0.5669 | 0.5706 | 0.5661 |
| MLP | 0.5690 | 0.5640 | 0.5670 |
| Random Forest | 0.5314 | 0.4166 | 0.4434 |

Best test model: **SVM RBF**.

## Usage

```python
import joblib
from sentence_transformers import SentenceTransformer

payload = joblib.load("best_model.joblib")
model = payload["model"]
metadata = payload["metadata"]
classes = metadata["label_classes"]

embedder = SentenceTransformer(metadata["embedding_model"])
text = "The final scene reveals that the detective was the killer all along."
X = embedder.encode([text], convert_to_numpy=True, normalize_embeddings=True)
label_id = int(model.predict(X)[0])
print(classes[label_id])
```

## Data

The training data was built from IMDb reviews and GPT-generated synthetic review snippets. GPT was also used to assign Mild/Major severity labels for IMDb spoiler reviews. A manual quality check of 100 sampled Mild/Major labels found 93% exact agreement.

## Limitations

Spoiler severity is subjective, especially between Mild and Major. Synthetic examples can also differ stylistically from real user reviews, so results should be interpreted as a course-project prototype rather than a production moderation system.