I had Gemini train its own replacement for $9

I decided to scrape Reddit threads about knives, using the Gemini 3.1 Pro model for named-entity recognition. This cost me $9 for 4,290 comments, which translates to a cost of $0.0021 per comment. After training a smaller model, GLiNER, on these labels, I found that upon running it on my own machine, the model matched Gemini's labels 83% of the time on unseen comments.

The model is capable of identifying product names with full accuracy and categorizes steel types correctly, such as recognizing that carbon steel is a category rather than a specific steel type. The entire project cost less than lunch - just $9 for labels and $2.50 for GPU usage. The main challenge was identifying a tensor filled with ones that was incorrectly used as a mask, causing the model to struggle in identifying spans.