HF engineer trains Qwen3.5-9B with GRPO and GEPA to code Blender 3D rooms, GLM-5.3-Flash as judge
mervenoyann · x · 2026-09-07
Hugging Face engineer Merve Noyann open-sourced a research project built on TRL and OpenEnv that trains Qwen3.5-9B to write bpy code via tool calls, generating complete low-poly isometric room dioramas from blank Blender geometry, with GLM-5.3-Flash serving as judge.
Key details:
- 256 deterministic training prompts across four room families (studio, bedroom, kitchen, living room) plus a 64-prompt held-out eval set
- The policy gets four tools: executepython, inspectscene, renderscene (text metadata only) and finish
- Two-phase plan: run GEPA to iterate on rewards first (early runs produced no renders; current ones render but lack semantic quality), then 200 GRPO optimizer steps
- The first HF Job is evaluation-only and does not modify the base model
Code, scripts, Dockerfile and docs are public in the merve/blender-grpo-gepa HF bucket, with live rollouts viewable online.
Related event: HF engineer trains Qwen3.5 to build 3D rooms in Blender with GRPO and GEPA(2 posts)→
More from coding & agent
- Codex adds experimental cross-context memory: notes and history search via config flag — reach_vb · 2026-09-07
- mitsuhiko's 'slop factory' autonomously decides to implement braces and lexical scoping — mitsuhiko · 2026-09-07
- Maintainer of 100+ Open-Source Repos Promises Same-Day Triage, Even for Agent-Filed Issues — doodlestein · 2026-09-07
- The brutal truth of vibe coding: fix one bug, three more appear — ayushtweetshere · 2026-09-07
- When you ask AI to change a CSS class and it suggests rm -rf node_modules — ThePeterMick · 2026-09-07
- Agent Swarms Are Wildly Expensive: METR's HF Hack Probe Cost $400k in API Credits — natesiggard · 2026-09-07