Use
Serve the base model with this adapter in vLLM (version 0.32.0 was used to check it):
vllm serve Qwen/Qwen3.5-4B --max-model-len 10240 --enable-lora --max-lora-rank 32 \ --lora-modules isometric=najmulhasan/qwen3.5-4b-isometric-scenes-lora
Send requests to the model name isometric. Use the environment's system prompt, put the scene description in the user message, and turn thinking off with chat_template_kwargs: {"enable_thinking": false}. The model answers with one SVG document. To run a file of prompts, use training/generate.py from the GitHub repository.
The adapter is meant for research on RL post-training with rendered, judged outputs, and for drawing small animated scenes of places and objects from short English descriptions. It was not trained or evaluated for scenes with people or animals, for long or detailed specifications, for other languages, or for SVG that must meet accessibility or production requirements. Run the environment's code gates on its output before rendering it in a browser.
Training
The adapter was trained with prime-rl at commit 35fcb52, with verifiers at commit 540f40f. The loss was prime-rl's default at that commit, IPO, with the reward minus the group mean as the advantage. LoRA has rank 32 and alpha 64 and is applied to the attention, Gated DeltaNet and MLP projections of the language model. The vision tower is frozen and unused, since all inputs are text. The learning rate was 1e-5.
Training ran for 150 steps of 128 rollouts (16 prompts, 8 answers each), drawn from the 1,600 training prompts, at temperature 1.0 with up to 8,192 tokens per answer. It used 2 × H100 GPUs, one for training and one for inference.
The reward first runs code checks: the answer must be one valid, safe SVG that renders to a non-blank frame. A code check for motion and the training judge's yes/no answers about the rendered frames are then averaged, and a figure seen by the judge scores 0.
A checkpoint was saved every 15 steps, and the one with the best validation reward under the training judge was selected: step 150, the last, at 0.813 (next best 0.773 at step 120, on 88 validation prompts with 4 answers each).
Evaluation
Both models answered every test prompt 4 times, with seeds 0 to 3, at temperature 1.0 with up to 8,192 tokens and thinking off. The evaluation judge, google/gemini-3.1-pro-preview, a preview model that was not used in training, scored the answers at temperature 0 with up to 8,192 tokens. Differences have paired bootstrap intervals over prompts, shares have cluster bootstrap intervals over prompts, and the blind comparison has a Wilson interval.
Table with columns: Prompts, Base, RL, Difference (95% CI)| Prompts | Base | RL | Difference (95% CI) |
|---|
| Test (240) | 0.078 | 0.489 | +0.411 [0.387, 0.435] |
Held-out category, space (one category, 120) | 0.082 | 0.561 | +0.479 [0.448, 0.511] |
The judge request failed for 2 of the 960 RL answers on test. Scoring them as 0 instead of leaving them out gives +0.409 [0.385, 0.432].
In a blind pairwise comparison by the author on 50 test prompts, the RL answer was preferred 40 times and the base answer 6 times, with 4 ties (win rate 0.87 excluding ties, Wilson 95% CI [0.74, 0.94]). On the 30 pairs where both answers reached the judge, the counts were 22, 5 and 3 (0.81 [0.63, 0.92]). The RL outputs have a recognisable style, so the blinding is weak.
The gain is larger under the training judge, google/gemma-3-27b-it. On the same pairs, Gemma scores the base model at 0.25 and the RL model at 0.81 on test, against 0.08 and 0.49 under Gemini, so its gain exceeds Gemini's by 0.15 [0.12, 0.18] on test and by 0.09 [0.04, 0.14] on space. Gemma is more lenient overall, and part of the larger gain may come from optimising against it.
On the test prompts, valid SVGs rose from 54% to 97% of answers, and animated ones from 25% to 91%. Of the answers that reached the judge, which were valid, rendered and not blank (base 510, RL 926), the scene was recognisable in 46%, up from 6%, and 10% were judged isometric by Gemini, up from 2%.
On distractor prompts, which add one person, animal or face figure to the scene, the RL model did not clearly draw the figure less often. On test, 27 of 160 base answers (17% [11%, 23%]) and 29 of 159 RL answers (18% [10%, 27%]) were judged to show one; on space, 11 of 80 and 8 of 80. Among answers that reached the judge, the share fell from 35% to 19%, because about the same number showed a figure while twice as many RL answers reached the judge.
This repository also holds:
final_eval/: every answer of both models, with its scores;
selection.json: the checkpoint selection;
versions.txt: the software versions and GPUs of the run.
The analysis/ folder in the GitHub repository reproduces the evaluation numbers from them.
Limitations
- Few answers are isometric: 10% of the RL model's answers that reached the judge on test, up from 2%.
- Neither judge was checked against human ratings. The evaluation judge is a preview API model, so rescoring later may give different scores, and the gain is larger under the training judge, so part of it may be specific to that judge.
- The RL model did not clearly learn to leave out the figure on distractor prompts: on test, 27 base and 29 RL answers were judged to show one.
- Evidence of generalisation is limited: the held-out set is one category, and all prompts come from the same generator.
- There was one training run and one baseline, the base model, and the blind comparison had one rater, the author.
Citation
@misc{hasan2026isometricscenes, author = {Najmul Hasan}, title = {Isometric Scenes: RL post-training Qwen3.5-4B to write animated isometric SVG}, year = {2026}, url = {https://github.com/najmulhasan-code/isometric-scenes}}