r2s-smoke-v9-w16-77dbfc7-30dc1c5d-47646604
completed —
Hypothesis
Trained on the 2,681 family-disjoint training icons with metric coordinates, path-major order, online exact augmentation and compositions, the 9.0M-parameter render-to-program model transcribes held-out validation renders better than retrieving the closest training icon. Pass requires all three: (1) mean 72 px pixel error below the nearest-training-icon baseline with the paired 95% interval on the reduction excluding zero; (2) CLIP top-1 on Gate N's 32 validation icons at or above OmniSVG 4B zero-shot, 0.609; (3) median single-icon latency on the RTX 4080 under 500 ms.
Visual output
Written result
Hypothesis. Trained on the 2,681 family-disjoint training icons with metric coordinates, path-major order, online exact augmentation and compositions, the 9.0M-parameter render-to-program model transcribes held-out validation renders better than retrieving the closest training icon. Pass requires all three: (1) mean 72 px pixel error below the nearest-training-icon baseline with the paired 95% interval on the reduction excluding zero; (2) CLIP top-1 on Gate N's 32 validation icons at or above OmniSVG 4B zero-shot, 0.609; (3) median single-icon latency on the RTX 4080 under 500 ms.
Result
| measure | value |
|---|---|
| icons evaluated | 32 (primary/validation) |
| model pixel error, mean [95% CI] | 0.1632 [0.1398, 0.1896] |
| model pixel error, median | 0.1326 |
| exact program rate | 0.000 |
| rendered rate | 1.000 |
| blank canvas pixel error | 0.1627 |
| nearest training icon pixel error | 0.0943 [0.0816, 0.1074] |
| error reduction vs nearest icon | -0.0689 [-0.0965, -0.0433] |
| icons where model beats nearest icon | 1 |
| CLIP top-1, model greedy (8 Gate N icons) | 0.000 |
| CLIP top-1, nearest training icon | 0.375 |
| CLIP top-1, OmniSVG 4B zero-shot (Gate N) | 0.609 |
Resources
| measure | value |
|---|---|
| device | NVIDIA H100 |
| parameters | 9,025,826 |
| train seconds | 121 |
| peak VRAM GiB | 4.38838529586792 |
| latency ms/icon, batch 1, median | 3183.0 |
| latency ms/icon, batch 1, p95 | 3186.0 |
| throughput ms/icon, batch 32 | 168.5 |
| model calls per icon (decoding) | 820.0 |
| torch / CUDA | 2.14.1+cu130 / 13.0 |
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
- running2026-10-13T17:41:52Z
- completed2026-10-13T17:46:20Z
Run record
Verbatim from runs/r2s-smoke-v9-w16-77dbfc7-30dc1c5d-47646604/run.yaml, the record committed before launch.
schema_version
2
run_id
r2s-smoke-v9-w16-77dbfc7-30dc1c5d-47646604
state
completed
planned_at
2026-10-13T17:41:52Z
hypothesis
Trained on the 2,681 family-disjoint training icons with metric coordinates, path-major order, online exact augmentation and compositions, the 9.0M-parameter render-to-program model transcribes held-out validation renders better than retrieving the closest training icon. Pass requires all three: (1) mean 72 px pixel error below the nearest-training-icon baseline with the paired 95% interval on the reduction excluding zero; (2) CLIP top-1 on Gate N's 32 validation icons at or above OmniSVG 4B zero-shot, 0.609; (3) median single-icon latency on the RTX 4080 under 500 ms.
parent_run
r2s-full-v9-colour
git_commit
77dbfc75cac8662c541b1d75887e02114702b7b8
dirty_patch_sha256
config
configs/render2svg/smoke-v9-w16.yaml
config_sha256
30dc1c5d3f515374…config_resolved
model
image_size
144
d_model
256
heads
8
encoder_layers
2
decoder_layers
6
feedforward
1024
dropout
0.1
metric
true
fourier
10
order
path
training
steps
1200
batch_size
32
learning_rate
0.0005
weight_decay
0.01
warmup_steps
100
seed
7001
eval_every
400
eval_icons
32
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
16
extra_training
bf16
true
final_eval_icons
32
clip_icons
8
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
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/smoke-v9-w16.yaml
outputs
run_dir
runs/r2s-smoke-v9-w16-77dbfc7-30dc1c5d-47646604
checkpoints
/data/fs201459/cp98556/predictable/cache/runs/r2s-smoke-v9-w16-77dbfc7-30dc1c5d-47646604
notes
completed_at
2026-10-13T17:46:20Z
resource
device
NVIDIA H100
torch_version
2.14.1+cu130
cuda_version
13.0
python_version
3.12.14
peak_vram_gib
4.38838529586792
train_seconds
121.19013075297698
inference_ms_per_icon
3182.9965619836003
inference_p95_ms_per_icon
3185.9754879260436
icons_per_second
0.31416936227440145
batched_ms_per_icon
168.5207129085029
batched_icons_per_call
32
excludes
rendering the output SVG to pixels
baselines
blank_canvas_pixel_error
0.16265273559838533
nearest_training_icon_pixel_error
0.09432368678972125
gate_n_omnisvg_zero_shot_top1
0.609375
nearest_training_icon_clip_top1
0.375
result
icons
32
model_pixel_error
0.16320836928207427, 0.13980680825479794, 0.1896097112330608
model_pixel_error_median
0.13258177787065506
exact_program_rate
0.0
rendered_rate
1.0
blank_pixel_error
0.16265273559838533
selected_step
1200
nearest_training_icon_pixel_error
0.09432368678972125, 0.08160671039076987, 0.10738429491466377
model_minus_baseline_error_reduction
-0.06888468249235302, -0.0964836315222783, -0.043297941182390776
icons_model_beats_baseline
1
clip_retrieval
model_greedy
top1_rate
0.0
top5_rate
0.625
mean_rank
4.75
clip_to_reference_mean
0.7855054810643196
chance_top1
0.125
icons
8
nearest_training_icon
top1_rate
0.375
top5_rate
1.0
mean_rank
2.125
clip_to_reference_mean
0.8378832936286926
chance_top1
0.125
icons
8
gate_n_omnisvg_zero_shot_top1
0.609375
gate_n_omnisvg_best_of_12_top1
0.828125
model_calls_per_icon
820