I had Gemini train its own replacement for $9 (www.petervijeh.com)

In a recent project, a developer utilized the Gemini 3.1 Pro model to label over 4,290 Reddit comments for brands, models, and materials associated with chef's knives, reducing ongoing costs significantly. Initially, using Gemini resulted in a growing expense of around $9 for labeling, as the paid API charged for each comment. The developer sought a more cost-effective solution using GLiNER, an open-source named entity recognition (NER) model. By training GLiNER on the labels generated by Gemini, they aimed to create a self-sufficient model that could run locally and maintain high accuracy. The end result was impressive, achieving an F1 score of 0.83 on a 225-comment validation set, which shows close alignment with Gemini's labeling. This approach not only cut costs down to a mere $2.50 in GPU usage after the initial labeling but also highlighted the potential of fine-tuning models for specific tasks without extensive resources. Although the development process involved troubleshooting numerous technical issues, such as correcting misconfigured training parameters and building the necessary input tensors, it ultimately proved an efficient method for scalable NER in specialized applications. This project exemplifies the growing trend of using open-source models for practical and economical AI solutions in niche domains.