r2s-cur-doodle-smoke-e9cea49-fb85b184-47646604
completed —
Hypothesis
Fine-tuned for 30k steps on renders degraded to look hand-drawn (probability 0.9), the final model turns degraded validation renders into programs closer to the clean icon than the parent model does, and closer than the degraded input itself (what a faithful tracer returns), while its error on clean validation renders rises by at most 0.010. Scored by doodle_eval on the sets declared in state/CURRENT.md (2026-10-31 16:40): clean, the training kinds and the held-out kinds pencil, hatched, affine, thumb.
Visual output
Written result
Hypothesis. Fine-tuned for 30k steps on renders degraded to look hand-drawn (probability 0.9), the final model turns degraded validation renders into programs closer to the clean icon than the parent model does, and closer than the degraded input itself (what a faithful tracer returns), while its error on clean validation renders rises by at most 0.010. Scored by doodle_eval on the sets declared in state/CURRENT.md (2026-10-31 16:40): clean, the training kinds and the held-out kinds pencil, hatched, affine, thumb.
Result
| measure | value |
|---|---|
| icons evaluated | 16 (primary/validation) |
| model pixel error, mean [95% CI] | 0.1633 [0.1138, 0.2157] |
| model pixel error, median | 0.1273 |
| exact program rate | 0.000 |
| rendered rate | 1.000 |
| blank canvas pixel error | 0.1829 |
| nearest training icon pixel error | 0.0775 [0.0569, 0.1018] |
| error reduction vs nearest icon | -0.0859 [-0.1470, -0.0305] |
| icons where model beats nearest icon | 5 |
| CLIP top-1, model greedy (4 Gate N icons) | 1.000 |
| CLIP top-1, nearest training icon | 0.250 |
| CLIP top-1, OmniSVG 4B zero-shot (Gate N) | 0.609 |
Resources
| measure | value |
|---|---|
| device | NVIDIA GeForce RTX 4080 |
| parameters | 9,025,826 |
| train seconds | 12 |
| peak VRAM GiB | 4.341804504394531 |
| latency ms/icon, batch 1, median | 2457.6 |
| latency ms/icon, batch 1, p95 | 2674.7 |
| throughput ms/icon, batch 16 | 177.5 |
| model calls per icon (decoding) | 932.0 |
| torch / CUDA | 2.14.0a0+4fdf77b940.nv26.08 / 13.4 |
Latency covers encoding and decoding to a validated token program; it excludes rasterising the SVG. Samples: samples.png, rows are reference, model, nearest training icon.
State transitions
- running2026-10-31T14:57:30Z
- completed2026-10-31T14:58:36Z
Run record
Verbatim from runs/r2s-cur-doodle-smoke-e9cea49-fb85b184-47646604/run.yaml, the record committed before launch.
schema_version
2
run_id
r2s-cur-doodle-smoke-e9cea49-fb85b184-47646604
state
completed
planned_at
2026-10-31T14:57:30Z
hypothesis
Fine-tuned for 30k steps on renders degraded to look hand-drawn (probability 0.9), the final model turns degraded validation renders into programs closer to the clean icon than the parent model does, and closer than the degraded input itself (what a faithful tracer returns), while its error on clean validation renders rises by at most 0.010. Scored by doodle_eval on the sets declared in state/CURRENT.md (2026-10-31 16:40): clean, the training kinds and the held-out kinds pencil, hatched, affine, thumb.
parent_run
r2s-cur-d02-s120k-seed2-53b3108-46294452-47646604
git_commit
e9cea49eac774e92bdc6703009317149df4c6876
dirty_patch_sha256
9fb968842c9ce601…config
configs/render2svg/cur-doodle-smoke.yaml
config_sha256
fb85b1841c508939…config_resolved
model
image_size
144
d_model
256
heads
8
encoder_layers
2
decoder_layers
6
feedforward
1024
dropout
0.2
metric
true
fourier
10
order
path
sources
0
training
steps
60
batch_size
32
learning_rate
0.0003
weight_decay
0.01
warmup_steps
500
seed
7001
eval_every
30
eval_icons
8
train_icons
None
evaluate_on_train
false
augment_variants
0
augment_seed
9001
augment_mirror
0.5
augment_max_shift
48
augment_colour
0.1
augment_online
true
compose_probability
0.5
compose_parts
4
augment_original
0.1
loader_workers
4
extra_training
degrade_probability
0.9
degrade_kinds
lineart, sloppy, flat
degrade_eval_seed
0
init_run
r2s-cur-d02-s120k-seed2-53b3108-46294452-47646604
init_sha256
38daeca4dd27fb4c…bf16
true
final_eval_icons
16
clip_icons
4
dataset
pilot_config
configs/learning/openmoji-g1-geometric-gate-v16.yaml
cache_sha256
476466042da98d72…train_icons
2681
augmented_variants
0
extra_training
None
augmented_sha256
None
evaluated_on
primary/validation
evaluation_inputs
renders degraded by doodle.degrade_split, kinds ['lineart', 'sloppy', 'flat'], seed 0; scored against the clean targets
init
run
r2s-cur-d02-s120k-seed2-53b3108-46294452-47646604
checkpoint
/home/dev/.cache/mojidiff/runs/r2s-cur-d02-s120k-seed2-53b3108-46294452-47646604/best.pt
sha256
38daeca4dd27fb4c…step
64000
seed
7001
determinism
seeded; cuDNN and SDPA kernels not forced deterministic
model_parameters
9025826
command
.venv/bin/python -m mojidiff.learning.render2svg --config configs/render2svg/cur-doodle-smoke.yaml
outputs
run_dir
runs/r2s-cur-doodle-smoke-e9cea49-fb85b184-47646604
checkpoints
/home/dev/.cache/mojidiff/runs/r2s-cur-doodle-smoke-e9cea49-fb85b184-47646604
notes
completed_at
2026-10-31T14:58:36Z
resource
device
NVIDIA GeForce RTX 4080
torch_version
2.14.0a0+4fdf77b940.nv26.08
cuda_version
13.4
python_version
3.12.3
peak_vram_gib
4.341804504394531
train_seconds
11.944061252987012
inference_ms_per_icon
2457.6142164878547
inference_p95_ms_per_icon
2674.7344110626727
icons_per_second
0.40689868787831446
batched_ms_per_icon
177.4943474374595
batched_icons_per_call
16
excludes
rendering the output SVG to pixels
baselines
blank_canvas_pixel_error
0.18288448872044683
nearest_training_icon_pixel_error
0.07745827548205853
gate_n_omnisvg_zero_shot_top1
0.609375
nearest_training_icon_clip_top1
0.25
result
icons
16
model_pixel_error
0.16331286635249853, 0.11380763738416136, 0.21574237653403544
model_pixel_error_median
0.12734558060765266
exact_program_rate
0.0
rendered_rate
1.0
blank_pixel_error
0.18288448872044683
selected_step
30
nearest_training_icon_pixel_error
0.07745827548205853, 0.05690388170769438, 0.10184130002744496
model_minus_baseline_error_reduction
-0.08585459087044, -0.1469769281859044, -0.030474699172191304
icons_model_beats_baseline
5
clip_retrieval
model_greedy
top1_rate
1.0
top5_rate
1.0
mean_rank
1.0
clip_to_reference_mean
0.9232548177242279
chance_top1
0.25
icons
4
nearest_training_icon
top1_rate
0.25
top5_rate
1.0
mean_rank
2.0
clip_to_reference_mean
0.8300648927688599
chance_top1
0.25
icons
4
gate_n_omnisvg_zero_shot_top1
0.609375
gate_n_omnisvg_best_of_12_top1
0.828125
model_calls_per_icon
932