I had Gemini train its own replacement for $9
3 hours ago
- Used Gemini to label 4,290 Reddit comments for knife brands, models, and steels at a cost of $9.
- Fine-tuned GLiNER large v2.5 on Gemini's labels to replace costly per-comment API calls with local inference.
- A critical bug: `words_mask` was mistakenly built as a binary mask instead of word indices, causing flat loss.
- Fixed the bug by emitting incremental word indices; model trained in 24 minutes on a Tesla T4.
- Achieved 0.83 F1 on a locked validation set, with material recall improved to 0.911 via per-class thresholds.
- Total project cost: $9 for labels and $2.50 for GPU time, enabling free local inference afterward.