35K Synthetic Examples for $16: Distilling Qwen3.8 on Hugging Face Jobs with datatrove

vanstriendaniel · x · 2026-08-18

Daniel van Strien of Hugging Face published the first guide on synthetic data generation with datatrove. Context: datatrove 0.10.0 adds a Jobs backend — pipelines run on HF Jobs with fan-out, multi-stage dependencies, retries and resume, no Slurm cluster needed — plus HF storage buckets as a DataFolder (hf://buckets/...) and preserved reasoning outputs from inference. Walkthrough: using Qwen/Qwen3.8-27B-FP8 as teacher, he generated tl;dr summaries over 36,000 model/dataset cards as training data for a small summarization model (card summaries are known to work better than raw cards as embedding input for search). The full 35,837-example dataset regenerated in one afternoon for $15.58, zero local GPU, with a calibration-first workflow and length-controllable outputs.

Related event: Hugging Face Generates 35K Model Card Summaries for Just $16(3 posts)→

Original post →

More from coding & agent

coding & agent channel →