# Teach Qwen to furnish a room Trained **Qwen3.5-9B with a rank-4 LoRA adapter** on 192 original procedural Blender programs. The full run, exact-weight reload and held-out evaluation completed on Compute. **Completed pilot GPU spend: $7.82.** This report covers the original seven finalized runs. The separately budgeted inference-only repair experiment is in progress and excluded from these totals. **Robustness limitation:** with Blender API hints, all eight adapter programs in a validation-only diagnostic failed with a list/tuple color-helper error. The original validation used batch 1, while this diagnostic used batch 8, so it does not isolate instructions from batching. The adapter remains limited to the tested setup. | Test policy | Rendered | Geometry valid | Token limit | |---|---:|---:|---:| | BASE | 0/32 | 0/32 | 3/32 | | SFT | 30/32 | 18/32 | 1/32 | One base test program was blocked by the frozen guard for a valid `bpy.ops.object.material_slot_add` call. Its execution/geometry is unmeasured; lifting that one restriction could add at most one base success. Another guard rejection imported `random`, contrary to the explicit task. See [evidence/test-guard-limitation.json](evidence/test-guard-limitation.json). These are 32 prompts from **four held-out layout groups**, each with eight correlated palette variants. Shared furniture primitives and a restricted Blender task limit the claim. This is a procedural-data control, not 192 independent Astra demonstrations, an open-ended design benchmark, or a comparison against Merve's different multi-turn GEPA/GRPO environment. Corrected validation: base **0/8 rendered, 0/8 geometry valid**; SFT **7/8 rendered, 5/8 geometry valid**. Required object names and axis-aligned bounds are diagnostics, not semantic, collision or aesthetic proof. There is no independent human/VLM preference score. A separate **post-hoc validation-only API-hint diagnostic** appended identical Blender 4.5 compatibility notes to both system prompts. Base rendered **1/8** and passed geometry on **0/8**; SFT rendered **0/8** and passed geometry on **0/8**. This diagnostic changed no model weights or primary test outputs. Its batch size also differed from original validation (8 versus 1), so it is not an isolated causal test of the API hints. See [evidence/api-hint-control-plan.json](evidence/api-hint-control-plan.json). ## Results and actual weights - [Published article](index.html), [all base/SFT comparisons](assets/results/gallery.html), [loss plot](assets/results/loss.png), [machine-readable results](assets/results/summary.json). - Actual adapter: `weights/sft/adapter/adapter_model.safetensors` (21,713,472 bytes). [Model card](MODEL_CARD.md). The ignored `weights/` directory is retained locally. - Portable full archive: `weights/sft-export.tar.gz`; final evaluation archive: `weights/test-export.tar.gz`. Their bytes and member hashes were verified. - [Execution and receipts](EXECUTION.md), [validation checks](VALIDATION.md), [research plan](PLAN.md), [weight verification](evidence/trained-weights-verification.json). Base/tokenizer revision: `c202236235762e1c871ad0ccb60c8ee5ba337b9a`. Adapter SHA-256: `a3ee9475724bfe2d84c264134c64f32e6b96452a8d5d594981921551e4360402`. All 496 tensors are finite bf16 values; all 248 LoRA-B tensors are nonzero. The selected checkpoint is step 72 of 72, after three epochs. Validation completion losses: **0.09904 → 0.02291 → 0.01696**. The training/validation loop took **37.16 minutes**; GPU receipts also include setup, model downloads, generation, rendering, and teardown. ## Reproduce Use Python 3.11 for local Blender preparation. The separate Compute SDK interpreter runs `submit.py` and `retrieve.py`; this experiment used the neighboring Compute checkout (SDK 0.1.17). Install/sign in through the [Compute quickstart](https://compute.cx/quickstart), or use that checkout's `.venv/bin/python`. `submit.py` uses its current client/quote interfaces. ```bash cd posts/blender-rooms uv venv --python 3.11 .venv uv pip install --python .venv/bin/python -r requirements.txt .venv/bin/python rooms.py --output data .venv/bin/python render.py --output artifacts .venv/bin/python -m pytest -q ``` All 256 trusted teacher programs were executed, rendered and geometry-checked. The model sees text and bpy targets; teacher images are quality-control assets. Completion/padding labels are masked. Targets are never truncated; observed training sequence lengths were 2,414–4,240 tokens, below the 6,144 limit. The deployed Compute agent had artifact-transfer bugs during this experiment. The included receiver uses immutable, authenticated, size-bounded archive slots. Start it in one terminal, then connect a temporary HTTPS tunnel in another: ```bash .venv/bin/python receiver.py --root transfer/received --key-file transfer/worker.key --port 8891 cloudflared tunnel --url http://127.0.0.1:8891 --protocol http2 ``` Save the printed public HTTPS origin in `transfer/base-url`. The private token is generated in `transfer/worker.key` and read directly by `submit.py`; keep that file private. The transfer directory is ignored by Git. Keep both processes running until all archives arrive, then stop them. Set `COMPUTE_PY` to the interpreter containing the Compute SDK. Run one stage at a time; use the returned run ID with `compute logs RUN_ID --follow` and `compute runs status RUN_ID`. The launcher refuses overlapping allocations and quotes beyond the remaining ledger budget. Use a fresh ledger for a new experiment, and verify a current GPU offer before each stage. ```bash COMPUTE_PY=/path/to/compute/.venv/bin/python "$COMPUTE_PY" submit.py smoke --gpu runpod/A100-PCIe-80GB --timeout 1800 --transfer-dir transfer --ledger evidence/my-budget.json --budget 25 "$COMPUTE_PY" retrieve.py SMOKE_RUN_ID --archive transfer/received/smoke.tar.gz --out weights/my-smoke "$COMPUTE_PY" submit.py train --gpu runpod/H100-SXM --timeout 10800 --transfer-dir transfer --ledger evidence/my-budget.json "$COMPUTE_PY" retrieve.py TRAIN_RUN_ID --archive transfer/received/sft.tar.gz --out weights/my-sft "$COMPUTE_PY" submit.py test --gpu runpod/H100-SXM --timeout 7200 --count 32 --training-run TRAIN_RUN_ID --transfer-dir transfer --ledger evidence/my-budget.json "$COMPUTE_PY" retrieve.py TEST_RUN_ID --archive transfer/received/test.tar.gz --out weights/my-test ``` Add `--plan-only` to a submission command to resolve source without network or spending. Full SFT uses learning rate 2e-4 with cosine decay, rank 4/alpha 8, batch 1 with eight accumulated steps, bf16 and gradient checkpointing. It selects by validation completion loss. Test generation uses identical left-padded batches of eight for both policies, greedy decoding, and 4,096 new tokens. The evaluator was corrected using validation outputs before the test: ordinary mesh/object operators are permitted and supplied lighting is fixed. The exact saved validation programs were replayed for both policies. Geometry thresholds and generated programs were unchanged. See the [correction record](evidence/evaluator-correction.json). Original and final source snapshots, hashes and archives are saved with each run. ```bash uv pip install --python .venv/bin/python -r requirements-report.txt .venv/bin/python report.py --training weights/my-sft --test weights/my-test --output assets/my-results ``` ## Use the adapter On a suitable NVIDIA GPU, load `Qwen/Qwen3.5-9B` at the pinned revision with `AutoModelForImageTextToText` in bf16, then `PeftModel.from_pretrained(base, "weights/sft/adapter")`. Use the pinned tokenizer, `rooms.SYSTEM` and the same non-thinking chat template. `inference.batch_tokens` returns code tokens without executing them; `evaluate.py` is a complete loading example. This adapter was evaluated on the four procedural room families only. ## Review and next experiment Share only [the anonymous review folder](assets/results/review/index.html) with reviewers before showing labeled results. It contains shuffled A/B images, the brief, and a local JSON export of their own choices. The identity key remains separate. No preferences have been filled in automatically. Because the primary base policy produced no completed renders, those pairs do not establish a fine-grained aesthetic advantage. The optional control can be reproduced with `submit.py control --count 8` and the same transfer/training arguments. It uses the receiver's `probe` slot; preserve any earlier throughput probe before that run. The predeclared 80% validation-geometry gate was not met (5/8). GRPO/GEPA was therefore not started. The next useful experiment is richer, independently authored layouts and a repair/geometry curriculum, followed by independent visual review. The completed pilot left **$17.18** before the separate follow-up. Generated programs are untrusted. The AST filter is not an OS sandbox; execute model code only on disposable workers without account secrets or sensitive mounts. Local preparation/tests execute this project's trusted demonstrations. Article: https://letsusecompute.com/posts/blender-rooms/ [Shareable GIF, MP4, and post copy](SOCIAL.md). No model was uploaded to Hub. The follow-up code is included for inspection; this article reports only the completed pilot. Private payloads, credentials, raw account records, and model archives are excluded from publication.