Prompt Engineering Experiments

Hands-on experiments showing how prompt phrasing changes the output of pretrained language models, using Hugging Face Transformers in Google Colab.

What this covers

Text generation (GPT-2) — the same subject is prompted several different ways to show how framing, specificity, and instruction style change the continuation the model produces.

Summarization (BART) — facebook/bart-large-cnn is used to summarize an article under different instructions, including a length-constrained "explain it to a child" variant, to compare how constraints affect the result.

Each experiment is followed by a short written analysis of why the outputs differ.

Stack

• Hugging Face Transformers (pipeline API)
• gpt2 for generation, facebook/bart-large-cnn for summarization
• PyTorch (as the Transformers backend)
• Google Colab

Running it

Open the notebook in Google Colab and run the cells top to bottom. No API key is needed — both models are downloaded from the Hugging Face Hub at runtime. A GPU runtime is faster but not required.

Notes

Model outputs are non-deterministic, so re-running the generation cells will not reproduce the committed outputs exactly. That variability is part of the point.