Cultural Sarcasm Detector (CSD)

A multi-task deep learning model that detects sarcasm in text and classifies the cultural domain it belongs to.

Model

• Backbone: DistilBERT (distilbert-base-uncased)
• Tasks: Sarcasm detection (score 0–1) + Domain classification (7 categories)
• Training: Mixed precision FP16 on NVIDIA RTX 3050

Datasets

• Sarcasm Headlines Dataset v2 — Rishabh Misra (Kaggle)
• Reddit Sarcasm Corpus — danofer (Kaggle)

Download both datasets and place them in data/raw/ before running.

Setup

conda create -n csd python=3.10 -y
conda activate csd
conda install pytorch==2.1.2 torchvision torchaudio pytorch-cuda=12.1 -c pytorch -c nvidia -y
pip install transformers==4.37.0 datasets==2.16.1 scikit-learn==1.4.0 pandas==2.1.4 numpy==1.26.3 tqdm==4.66.1 matplotlib==3.8.2 seaborn==0.13.1 accelerate==0.26.1

Usage

Train:

python src/train.py

Evaluate:

python src/evaluate.py

Interactive demo:

python src/inference.py

Results

Metric	Score
Accuracy	0.7715
F1 Score	0.7715
AUC	0.86

Project Structure

csd/
├── config.py          # All hyperparameters and settings
├── src/
│   ├── data_loader.py # Dataset loading and preprocessing
│   ├── preprocess.py  # Tokenization and DataLoader creation
│   ├── model.py       # DistilBERT + dual classification heads
│   ├── train.py       # Training loop with FP16 and scheduler
│   ├── evaluate.py    # Metrics, plots, and evaluation
│   └── inference.py   # Inference and interactive demo
├── data/raw/          # Place datasets here (not tracked by git)
├── models/            # Saved model weights (not tracked by git)
└── logs/              # Training plots and history

Pretrained Model

The trained model weights (models/csd_model.pt) are stored via Git LFS. They are downloaded automatically when you clone the repo:

git clone https://github.com/arnavgoel17/Cultural-Sarcasm-Detector.git

No retraining needed — run inference immediately after cloning:

python src/inference.py

Domain Classes

politics · entertainment · technology · sports · lifestyle · humor · general

Known Limitations

• Domain classification defaults toward general due to class imbalance in Reddit training data (74% of subreddits mapped to general). Sarcasm detection is unaffected — trained on a balanced 50/50 split.