sergiopaniego/pelican-svg-drawings

Pelican SVG: frontier model drawings, scored

139 SVGs produced by seven frontier models answering Simon Willison's prompt, "generate an SVG of a pelican riding a bicycle", each with the score the pelican_svg_env OpenEnv environment gave it. Four of the seven are open weights and three are closed.

Simon has run that prompt against nearly every model release since early 2025, but the results live as embedded images across 129 blog posts and scattered gists. This dataset exists because there was no machine-readable version: the SVG source, the rendered PNG and a reproducible score, in one place.

This is not a leaderboard. Simon's own position is that "the correlation between pelican performance and actual model quality has been mostly severed now" and "don't go using pelicans to compare models". The numbers below agree with him, and the section on the geometry layer explains why.

Configs

config	rows	what it is
default	139	the canonical prompt, 7 models, 20 samples each
coverage	30	one sample of each of the 30 subject-vehicle combinations

What is in a row

field	meaning
model, task, sample	which model, which prompt, which of its 20 samples
access, provider	open or closed weights, and which API served it
reward	0.35 x structure + 0.65 x semantic, zero if the gate rejected it
structure_score	fraction of 7 deterministic geometry checks passed
semantic_score	vision judge: blind caption plus a binary feature checklist
gate_passed, violations	whether it was admitted, and why not
judge_model	which judge scored it. The judge dominates the score, so a number without this is not reproducible
judge_blind_caption	what the judge said the drawing was, asked with no mention of the task
svg, png	the source and its render

Where the models landed

model	access	n	mean reward	structure	semantic	perfect
gpt-5.6-sol	closed	20	0.959	1.000	0.938	13
moonshotai/Kimi-K3	open	19	0.884	0.947	0.849	11
zai-org/GLM-5.2	open	20	0.878	1.000	0.812	7
Qwen/Qwen3.6-27B	open	20	0.870	1.000	0.800	6
claude-opus-5	closed	20	0.868	1.000	0.797	6
claude-fable-5	closed	20	0.844	1.000	0.760	1
nvidia/NVIDIA-Nemotron-3-Ultra-550B-A55B-NVFP4	open	20	0.814	0.986	0.722	3

The three open-weight models were sampled at temperature 0.7 through Hugging Face Inference Providers. Both closed vendors reject a temperature on these models, so those runs use whatever their API does by default: compare the means with that in mind and do not compare the spreads at all. Neither closed API returns the reasoning it bills for, which is why reasoning_chars is 0 for them rather than missing.

The geometry layer is saturated

136 of the 139 samples scored exactly 1.000 on structure. Five of the seven models scored 1.000 on every one of their 20 samples. At frontier level the deterministic layer is a formality and all of the remaining signal is in the judge. It still earns its place for training a small model and for refusing the two cheats, but reading it as a quality measure is a mistake.

The finding worth having

judge_blind_caption is the interesting column. The judge is shown each drawing cold and asked what it is, so its answer cannot be inflated by a leading question. Split by model, each one has a signature near-miss and they are not the same bird:

model	judge says "pelican"	otherwise, most often
gpt-5.6-sol	13/20	bird 4
moonshotai/Kimi-K3	11/19	stork 7
zai-org/GLM-5.2	9/20	goose 5
Qwen/Qwen3.6-27B	7/20	stork 10
NVIDIA-Nemotron-3-Ultra	7/20	bird 7
claude-opus-5	7/20	goose 6
claude-fable-5	1/20	duck 8

Frontier models reliably draw a bird on a bicycle, but only 37% of the time is it identifiably a pelican. Ask "is this a pelican?" and you get a yes; ask what it is and you get a stork, or a duck.

A correction, kept visible

12 rows carry a structure_score_before_use_fix field. Those were originally scored by a shape walker that descended into <defs>, whose content is never painted, and ignored <use>, which is what paints it. A model that wrote its wheel once and placed it twice had the template counted as a single wheel at the origin and both real wheels discarded. GLM-5.2 scored 0.000 on a drawing this dataset's own judge caption calls "a cartoon pelican is riding a bicycle".

Every corrected row moved upward, and it changed the ordering of the table above. Before the fix the two mid-table open models outranked both Claude models; after it the middle four sit within 0.034 of each other, which at n=20 is no separation at all.

The coverage config

One sample of each of the 30 combinations, all admitted by the gate, none crashed. Reward ranged 0.362 to 1.000 with a median of 0.838. Read a single sample as "this combination is wired up", never as a result about the combination: there is no variance in one draw to report.

Reproducing it

python examples/pelican_svg_eval.py \
    --env-url https://sergiopaniego-pelican-svg-env.hf.space \
    --samples 20 --save-svgs ./svgs

Generation ran through the providers named above; scoring ran inside the environment, never in the harness. Kimi K3 is absent on purpose: released two days before this run, it failed 19 of 20 samples on a 504 at exactly the router's 120-second gateway limit and then hit rate limits.

Environment and scoring code: OpenEnv.