Fine-tuning a 194M GLiNER2 dataset tagger on HF Jobs costs $1.50, lifting accuracy 10% to 69%

iamrobotbear · x · 2026-09-25

Daniel van Strien trained a Jev-style classifier on Hugging Face Jobs for about $1.50.

• Model: a 194M GLiNER2 that suggests task tags for any Hub dataset from column names and the first row, returning labels with probabilities
• Results: zero-shot, GLiNER2's first suggestion matched an owner's tag 10% of the time; after 17 minutes of fine-tuning, 69% (owner tags are noisy, so some "wrong" answers are tags the owner left out)
• Deployment: the fine-tuned model runs on a free CPU in about a second
• Open recipe: one hf jobs command trains the same kind of model on your own labels; the README example (book titles) runs in 2 minutes for $0.02
• Demo and recipe are public

Original post →

More from Infra

• Splash on a 40-core M5 Max: +20% decode speed by running tune-kernels on your own chip — SeveralViolins · 2026-09-25
• 100-agent doc review fan-out finishes in 10 minutes for $5.75, open-source harness holds — Pitiful-Surround-285 · 2026-09-25
• Nvidia's UHP laser demand ramps for December quarter as Spectrum-6 CPO beats plan, says Lumentum — firstadopter · 2026-09-25
• Lambda's Stephen Balaban on 14 years, 5 pivots, and debunking data center myths — gharik · 2026-09-25
• Jim Chanos: Much of the financial economy now depends on the AI buildout — GaryMarcus · 2026-09-25
• Gemini 3.8 Flash benchmarked so fast the engineer thought his Slurm job crashed — _lewtun · 2026-09-25

Infra channel →