PretrainGym open-sources benchmark for testing frontier models at pretraining data curation
xeophon · x · 2026-11-03
PretrainGym, built during the PrimeIntellect RL Residency, is now open source. The project tackles a less-studied piece of recursive self-improvement: how good frontier models actually are at curating pretraining data. It is built entirely on PrimeIntellect's verifiers library, which handles much of the engineering pain of agent evaluations.
More from coding & agent
- Skill++ paper: routing-aware skill refinement lifts agent routing from 36.4% to 76.3% — dair_ai · 2026-11-03
- Herald: open-source multi-channel notification system handling 4.3M notifications/day — damnGruz · 2026-11-03
- Dev builds 3D live desktop showing Claude Code and Codex agents, art fully generated by Claude Code — xiaohu · 2026-11-03
- Dev mocks pre-AI SWE nostalgia: nobody misses fixing CI and merge conflicts by hand — tdhopper · 2026-11-03
- Claude Code prompt hunts down vulnerable Chromium builds across your Mac in 10 minutes — Kyrannio · 2026-11-03
- "Claude Code can build it for free" — automation freelancer explains why judgment now beats code — Salman94157 · 2026-11-03