LLM Non-Determinism Experiment

This project demonstrates that Large Language Models (LLMs) produce probabilistic outputs.

Objective

To show that the same prompt can generate different responses when sampling parameters like temperature are used.

Models Used

• GPT-2
• DistilGPT2
• Phi-2

Tools

• Python
• Hugging Face Transformers

Experiment Steps

1. Send the same prompt to the LLM multiple times.
2. Store responses in a list.
3. Compare outputs.
4. Change temperature parameter to observe variation.

Example Prompt

Explain machine learning in simple words.

Observation

Different outputs are generated for the same prompt due to probabilistic token sampling.

Installation

pip install -r requirements.txt python llm_test.py