GenAI Tools: CIFAR-10 Image Classification with a Vision-Language Model (gemma3:4b)

This repo samples 100 CIFAR-10 images (10 per class), sends each image to an OpenAI‑compatible endpoint at https://ai.sooners.us using the gemma3:4b model, and logs predictions to compute accuracy and a confusion matrix.

1. Setup

Requirements

• Python 3.9+ recommended
• pip install -r requirements.txt

Create your SoonerAI account & API key

1. Visit https://ai.sooners.us and sign up with your OU email.
2. After logging in: Settings → Account → API Keys.
3. Create a new API key and copy it.

Create local env file (DO NOT commit)

Create ~/.soonerai.env and put the following code into the file. SOONERAI_API_KEY=your_key_here SOONERAI_BASE_URL=https://ai.sooners.us SOONERAI_MODEL=gemma3:4b

run the chatbot

python cifar10_classify.py

Analysis

Setup summary:

10 images per CIFAR-10.

Model: gemma3:4b via https://ai.sooners.us/api/chat/completions.

Temperature = 0.0 for deterministic output.

Failure modes:

Small objects or heavy background clutter. Low-contrast animal photos.

Conclusion

Explicit, short definitions yield the highest usable accuracy while maintaining valid single-word outputs.