Skip to content

README.md

Latest commit

 

History

History
70 lines (50 loc) · 2.45 KB

README.md

File metadata and controls

70 lines (50 loc) · 2.45 KB

Project: “Green AI Automated RAG Testing with Gemma”

Overview

We built an interactive retrieval-augmented generation (RAG) pipeline using the open-weight model Gemma 2-2b by Google and applied it to a standardized text and prompt set derived from the Apollo 11 lunar landing.

The goal:

  • evaluate summarisation, reasoning, retrieval, paraphrasing, and creative tasks in a controlled, reproducible way — logging both answer quality and local sustainability metrics (energy/carbon emissions) via CodeCarbon.

Model

Model ID: google/gemma-2-2b-it (Hugging Face)

Key attributes:

  • Open-weight decoder-only model trained by Google
  • Supports text-generation and conversational usage
  • Suitable for research, summarisation, reasoning, and retrieval tasks
  • Lightweight enough for deployment on modest compute resources

Model link: https://huggingface.co/google/gemma-2-2b-it

What We Did

  1. Created a source document (source.txt) using ~1,400 words of selected Wikipedia excerpts on Apollo 11.

  2. Defined a set of 21 standardised prompts spanning five categories: summarisation, reasoning, RAG (fact retrieval), paraphrasing, and creative generation.

  3. Built a document retrieval component using sentence-transformers to chunk the document and select top-k relevant chunks per query.

  4. Developed an interactive notebook workflow that:

    • Accepts a question at runtime
    • Runs RAG → Draft → Critic → Refiner cycles using Gemma
    • Tracks local CPU/GPU energy usage and CO₂ emissions with CodeCarbon
    • Logs each question, answer, timestamp, and emissions to a single append-only log file

  5. Logged runtime latency and emissions per query for performance and sustainability insights.

How to Use

  1. Clone the repository:

    git clone <YOUR_REPO_URL>
cd your_repo_folder

  2. Place your source.txt into ./data/.

  3. Add your Hugging Face API key in the config cell.

  4. Run the notebook setup cells, then use the interactive prompt cell to ask questions.

Why This Matters

  • Reproducibility — fixed source text and prompt set allow consistent evaluation across models.
  • Efficiency vs. Accuracy — emissions are logged alongside outputs to explore trade-offs between model performance and energy cost.
  • Accessibility — uses an open model and standard Python tools, making research on small language models feasible even on laptops.