The Last Thing You Ate Is What You Have to Name Him

I generally use pre-trained LLMs for my various little demo apps, but I decided to train one up for fun in my usual half-assed fashion.  Here's an image I asked the model to ID:

The uncanny result:

Overall:

In conclusion: 60% of the time, it works every time.

UPDATE: Training on a dataset 5x larger resulted in 100% confidence that the cute puppy is not, in fact, pizza.