tcjordan3/bert-base-spoiler-detection

BERT Base Spoiler Detection

Model Description

This model is a fine-tuned version of bert-base-uncased for detecting spoilers in movie and TV show reviews. It classifies reviews as either containing spoilers or being spoiler-free.

Developed by: Tyler Jordan Model type: Text Classification Language: English License: MIT Base model: bert-base-uncased

Intended Use

Primary Use Case

Automatically detect spoilers in user-generated movie and TV show reviews to warn readers before they encounter plot-revealing content.

Intended Users

• —Movie review platforms
• —Content moderation systems
• —Personal projects for filtering spoilers

Out-of-Scope Uses

• —Reviews in languages other than English
• —Non-entertainment content (news, academic papers, etc.)
• —Legal or medical content requiring high accuracy

Training Data

Dataset: IMDB Review Dataset by Enam Biswas (2021)

Preprocessing:

• —Sampled 200,000 balanced reviews (100k spoilers, 100k non-spoilers) from 5.5M total reviews
• —Train/Validation/Test split: 140k/30k/30k (70%/15%/15%)
• —Text cleaning: HTML tag removal, whitespace normalization
• —Minimum review length: 30 characters
• —Maximum sequence length: 512 tokens

Class Distribution:

• —Spoiler: 50%
• —Non-spoiler: 50%

Training Procedure

Training Hyperparameters

• —Optimizer: AdamW
• —Learning rate: 1e-5
• —Batch size: 32
• —Epochs: 5
• —Max sequence length: 512
• —Dropout: 0.3
• —Weight decay: 0.01
• —Warmup steps: 10% of total steps
• —Learning rate schedule: Linear warmup with decay

Training Hardware

• —GPU: NVIDIA T4 (Google Colab)
• —Training time: ~2-3 hours

Framework

• —PyTorch 2.5.1
• —Transformers 4.x
• —CUDA 12.1

Evaluation

Metrics

Evaluation Data

• —30,000 held-out reviews from the IMDB dataset
• —Balanced split (50% spoilers, 50% non-spoilers)