380M model curates 1M FineWeb pages in 19 min for ~$5, open recipe shared
vanstriendaniel · x · 2026-10-29
Developer Daniel van Strien trained a small specialist decision model for curating web-scale data, fine-tuned from UltraBERT-380M and cheap enough to run at scale.
- Processed 1M FineWeb pages in 19 minutes for $5 on Hugging Face Jobs
- The model flags 25% of pages as product pages; only 6% passed a simple "educational, not spam" filter, and 1 page in 1,020,000 was rated "outstanding"
- The model and the full recipe (data building, training, evaluation) are open-sourced and reproducible with a single command on HF Jobs
Related event: Open-Source 380M Model Curates 1M Webpages for $5(2 posts)→
More from Infra
- Ollama's two-week model drop: LFM3-1.7B targets edge agents with 256K context at ~1GB — ollama · 2026-10-30
- Meta's GenRecs Tokenization lifts ad conversions +0.35% with semantic vocabulary — Meta_Engineers · 2026-10-30
- Parallel Language Models paper: multi-token generation at AR quality, up to 5.7x speedup — aaron_lou · 2026-10-30
- Oracle investor day: 1GW AI site earns ~$80B contract value at ~28% ROI, 3.6-year payback — rohanpaul_ai · 2026-10-30
- Sam Altman: Local inference is a terrible allocation of scarce global compute — morqon · 2026-10-30
- Dual RTX 4060 Ti local LLM benchmarks: 35B MoE decodes at 122.5 tok/s — Rooneybuk · 2026-10-30