HF's merve shares GEPA agent-training rollouts built on TRL and OpenEnv
mervenoyann · x · 2026-09-07
Hugging Face engineer merve shares her agent-training project built on TRL and OpenEnv, rerunning experiments with GEPA after a few fixes. The rollout artifacts — low-poly room 3D scenes (.glb) in the blender-grpo-rollouts repo — are public on Hugging Face for others to try.
Related event: HF engineer trains Qwen3.5 to build 3D rooms in Blender with GRPO and GEPA(2 posts)→
More from coding & agent
- High-bandwidth error feedback to models matters more than raw model intelligence — akbirthko · 2026-09-07
- FrankenTerm brings FrankenTUI terminal interfaces to the browser via WebAssembly — doodlestein · 2026-09-07
- Do code quality norms built for human maintainers still hold as models approach superhuman coding? — burny_tech · 2026-09-07
- Astra ports Levin team's epiplexity estimator into live demo of evolving cellular automata — burny_tech · 2026-09-07
- Codex tip: let it create custom sections and rename threads for better relevance — jacob_posel · 2026-09-07
- Claude Code creator Boris Cherny: there's no one weird trick, iterate empirically — rohanpaul_ai · 2026-09-07