Dataset Builder agent: one prompt to keep ML training datasets fresh monthly
nlarusstone · x · 2026-11-06
Day 18 of nlarusstone's "an agent a day" series targets teams training or finetuning models for scientists, tackling three pain points: finding the right data, capturing the experimental context it was generated in, and keeping the dataset current.
The post includes a reusable sample prompt: create a Dataset Builder Agent that builds ML training datasets monthly; for existing datasets, find new data to add, check for new data that could form new datasets, and include experimental caveats/notes alongside the data. The pattern: hand data discovery, context annotation, and incremental maintenance to a single recurring agent.
More from coding & agent
- Dev essay: why a runaway intelligence explosion isn't the risk — 'early onset decoherence' is — round · 2026-11-06
- Sardine launches Know Your Human protocol to verify humans behind shopping agents — astrange · 2026-11-06
- Indie dev cuts 5,672 lines rebuilding trip planner's Days screen into one-drag editing — SeanOliver · 2026-11-06
- LangChain details optimizing browser use for deep agents; design agent fixes its own blog formatting — LangChain · 2026-11-06
- Baker Lab + NVIDIA unveil RFD4-Proteina: agents run protein design tool calls on GFP — AllThingsApx · 2026-11-06
- New Compound Writing skill PR lands on GitHub, pending testing before merge — kieranklaassen · 2026-11-06