A Builder Paid Gemini $9 to Label 4,290 Reddit Comments, Then Trained a 459M Model to Do It Locally
petervijeh.com / Hacker News (87pts, 43 comments)·low signal
Peter Vijeh had Gemini 3.1 Pro label 4,290 Reddit comments once for knife brands, models and steel types, computed character offsets programmatically, and fine-tuned GLiNER large v2.5 (459M parameters, DeBERTa-v3-large encoder) on a Tesla T4. Total cost was $9 in Gemini labels plus about $2.50 of GPU time across 10 training runs, yielding 0.83 F1 on held-out validation with 0.904 brand F1 and 0.911 material recall. The resulting model runs locally with no per-comment API cost, which is the whole point for anyone doing high-volume extraction.