I had Gemini train its own replacement for $9
A developer used Gemini 3.1 Pro to label 4,290 Reddit comments for $9, then fine-tuned GLiNER 459M to 0.83 F1 for product NER.
The author replaced per-comment Gemini 3.1 Pro API calls with a GLiNER large v2.5 (459M parameters) model fine-tuned on 4,290 Reddit comments that Gemini labeled for $9 via OpenRouter. Zero-shot GLiNER scored roughly 0.65 F1 against Gemini's labels; the fine-tuned model reached 0.83 F1 after 24 minutes on a Tesla T4, with about $2.50 of GPU cost. Key techniques included asking Gemini for exact substrings rather than character offsets, adding negative examples, and locking a 225-comment validation set; five of ten training runs failed on configuration and words_mask bugs.
- Gemini 3.1 Pro labeled 4,290 comments for $9, about $0.0021 per comment
- Fine-tuned GLiNER large v2.5 reached 0.83 F1 against Gemini's labels
- Zero-shot GLiNER baseline scored approximately 0.65 F1
- Labels were requested as substrings; offsets computed in code to avoid span errors
- Five of ten training runs failed on config and words_mask issues
- Vendors
- GoogleOpenRouterModal
- Products
- GLiNERGemini 3.1 Pro
- AI models
- Gemini 3.1 ProGLiNER large v2.5
From the original article
7 min read at petervijeh.comI like to cook, and somewhere along the way that turned into an obsession with high-end chef's knives. So I scrape the Reddit threads where people argue about them and pull out every brand, model and steel they mention, to see what is getting bought and argued about. Picking product names out of text is a job called named-entity recognition, and small models have done it for a decade.
By p-s-v · petervijeh.com
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.