😏 totally-not-sarcastic

Context-aware sarcasm and irony detection powered by RoBERTa.
Trained on 100k Reddit comments so you don't have to read them.

---

📊 Performance

Metric	Score
Accuracy	80.57%
F1	80.39%
Precision	81.14%
Recall	79.66%

---

🗂️ Repository Structure

totally-not-sarcastic/
├── colab_1_train.py      # Train RoBERTa and push to HuggingFace Hub
├── colab_2_dashboard.py  # Gradio dashboard — runs on HF Spaces or Colab
├── requirements.txt      # Dependencies for HF Spaces
└── README.md

---

🚀 Quick Start

Run the dashboard locally

pip install transformers gradio plotly lime torch
python colab_2_dashboard.py

Use the model directly

from transformers import pipeline

classifier = pipeline("text-classification", model="AK-Rahul/sarcasm-roberta")

# Without context
classifier("Oh absolutely, I love waiting 3 hours at the DMV.")

# With context — improves accuracy for conversational sarcasm
classifier("How was the flight? </s></s> Oh wonderful, only delayed by 4 hours.")

---

🧠 Model

Base: roberta-base (125M parameters)
Task: Binary classification — Sarcastic / Not Sarcastic
Context: Supports optional parent_comment as context for conversational input

Training Data

Dataset	Rows	Context
Reddit SARC (danofer)	~90,000	✅ parent_comment pairs
TweetEval Irony	~3,600	❌
News Headlines	~28,600	❌
Total (balanced)	107,058

Training Details

• 3 epochs · batch=32 · lr=2e-5 · fp16
• Label smoothing=0.05
• Embeddings frozen for epoch 1 (prevents catastrophic forgetting)
• Cosine LR decay · Early stopping (patience=2)
• Best checkpoint selected by validation loss

---

🗃️ Files

colab_1_train.py

Run this in Google Colab (T4 GPU) to train from scratch and push to HuggingFace Hub.

Requirements before running:

• danofer-sarcasm.zip in the root of your Google Drive
• HF_TOKEN added to Colab Secrets (write access)

Steps:

1. Runtime → Change runtime type → T4 GPU
2. Add HF_TOKEN to Colab Secrets
3. Paste entire file into one cell → Run
4. ~38 min training time

colab_2_dashboard.py

Gradio dashboard with three tabs — single prediction with LIME explanations, batch classification, and model stats.

To deploy on HuggingFace Spaces:

1. Create a new Space → SDK: Gradio
2. Upload this file as app.py
3. Upload requirements.txt
4. Set MODEL_REPO = "AK-Rahul/sarcasm-roberta" (already set)

To run in Colab:

• Paste into a cell and run — it will share a public Gradio link automatically

---

🖥️ Dashboard Features

• Sarcasm probability dial — semicircular gauge showing exact confidence
• Intensity bands — from ✅ Clearly Sincere to 🔥 Scorching Sarcasm
• LIME word attribution — highlights which words pushed the prediction
• Batch mode — classify multiple messages at once with shared context
• Export history — download all predictions as CSV
• Model stats tab — confusion matrix, radar chart, architecture details

---

📦 Requirements

transformers>=4.40
torch
gradio>=4.0
plotly
lime

---

📄 License

MIT — see LICENSE

---

🙏 Credits

• danofer/sarcasm — Reddit SARC dataset
• TweetEval — Irony benchmark
• raquiba/Sarcasm_News_Headline — News headlines
• RoBERTa — Base model by Meta AI