GPT-2 Language Model Exploration & Analysis — ShadowFox AIML Internship (Advanced Task)

This repository contains a complete Google Colab implementation of GPT-2.
The goal is to analyze how GPT-2 generates language, how well it understands context,
and how its behavior changes with temperature and domain-specific prompts.

---

Open in Colab

🔹 Project Summary

Item	Status
Large Language Model	GPT-2
Notebook Name	GPT2_analysis.ipynb
Framework	Hugging Face Transformers
Device	GPU (if enabled in Colab)

This notebook showcases: ✅ Text generation experiments
✅ Perplexity scoring
✅ Lexical diversity measurement
✅ Repetition analysis
✅ Visualizations (word cloud, token confidence, performance charts)
✅ Ethical considerations + conclusions

---

📌 Key Experiments

We tested GPT-2 on different prompts:

• Technology
• Story / narrative
• Medical domain
• Q&A format

And compared across multiple temperature settings:

Observed Effects:

Temperature ↑	Creativity ↑	Accuracy ↓	Hallucination ↑
✅ More surprising output	❌ Less factual

Domain Knowledge:

Domain	Performance
Story	⭐⭐⭐⭐
Tech	⭐⭐⭐
Medical	⭐⭐ (hallucination risk)

---

📊 Visual Results Included

• Perplexity vs Temperature Plot
• Distinct-1 Diversity Chart
• Token Confidence Curve
• Word Cloud of All Generated Text

These help demonstrate strengths and weaknesses in GPT-2 reasoning.

---

🔍 Research Questions

The notebook answers: 1️⃣ How does temperature impact text quality?
2️⃣ Does GPT-2 maintain context over longer sequences?
3️⃣ How bad is GPT-2 on domain-specific tasks (medical)?

Each question is supported with: ✔ Metrics
✔ Output examples
✔ Visualizations

---

⚠ Ethical Considerations

GPT-2:

• Produces confident but false information
• May generate biased/harmful text
• Not suitable for critical domains without safeguards

This is documented in the analysis section.

---

✅ Proof of Work Deliverables

Requirement	Provided
Colab Notebook	✅
Experiment Results	✅
Visualizations	✅
Ethical + Research insights	✅
Screenshots for submission	✅ Required separately
LinkedIn Proof of Work video	✅ Required separately

---

▶ How to Run This Notebook

1️⃣ Open GPT2_analysis.ipynb in Google Colab
2️⃣ Runtime → Change Runtime Type → GPU
3️⃣ Run all cells in order
All dependencies auto-install inside the notebook.

No local setup required.

---

📸 Output & Visual Results

Below are sample visualizations from the GPT-2 analysis:

✅ Word Cloud (Generated Text)

✅ Perplexity vs Temperature

✅ Token Confidence Visualization

✅ Example Generation Output

👤 Author

Ani
ShadowFox AIML Intern
Project Type: Advanced Level — Language Model Deployment & Evaluation

---

⭐ If you found this useful, star the repo!