r2s-full-v9-colour-989aea8-b26ad95e-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 | 339 (primary/validation) |
| model pixel error, mean [95% CI] | 0.0724 [0.0678, 0.0772] |
| model pixel error, median | 0.0675 |
| exact program rate | 0.021 |
| rendered rate | 1.000 |
| blank canvas pixel error | 0.1720 |
| nearest training icon pixel error | 0.0897 [0.0848, 0.0947] |
| error reduction vs nearest icon | 0.0173 [0.0131, 0.0214] |
| icons where model beats nearest icon | 234 |
| CLIP top-1, model greedy (32 Gate N icons) | 0.562 |
| CLIP top-1, nearest training icon | 0.344 |
| CLIP top-1, OmniSVG 4B zero-shot (Gate N) | 0.609 |
Resources
| measure | value |
|---|---|
| device | NVIDIA H100 |
| parameters | 9,025,826 |
| train seconds | 3341 |
| peak VRAM GiB | 4.38838529586792 |
| latency ms/icon, batch 1, median | 2940.1 |
| latency ms/icon, batch 1, p95 | 2958.9 |
| throughput ms/icon, batch 64 | 98.1 |
| model calls per icon (decoding) | 692.0 |
| 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
- running2026-10-13T17:51:23Z
- completed2026-10-13T18:51:09Z
Run record
Verbatim from runs/r2s-full-v9-colour-989aea8-b26ad95e-47646604/run.yaml, the record committed before launch.
schema_version
2
run_id
r2s-full-v9-colour-989aea8-b26ad95e-47646604
state
completed
planned_at
2026-10-13T17:51:23Z
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-v7-systems
git_commit
989aea8e4831d71e0462e79effc8b514161cc550
dirty_patch_sha256
config
configs/render2svg/full-v9-colour.yaml
config_sha256
b26ad95e18999bd4…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
60000
batch_size
32
learning_rate
0.0005
weight_decay
0.01
warmup_steps
1000
seed
7001
eval_every
2000
eval_icons
64
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
339
clip_icons
32
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/full-v9-colour.yaml
outputs
run_dir
runs/r2s-full-v9-colour-989aea8-b26ad95e-47646604
checkpoints
/data/fs201459/cp98556/predictable/cache/runs/r2s-full-v9-colour-989aea8-b26ad95e-47646604
notes
completed_at
2026-10-13T18:51:09Z
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
3340.9784624810563
inference_ms_per_icon
2940.117383550387
inference_p95_ms_per_icon
2958.8599409908056
icons_per_second
0.34012247456339095
batched_ms_per_icon
98.0689356874791
batched_icons_per_call
64
excludes
rendering the output SVG to pixels
baselines
blank_canvas_pixel_error
0.1720480552069557
nearest_training_icon_pixel_error
0.08973124639576185
gate_n_omnisvg_zero_shot_top1
0.609375
nearest_training_icon_clip_top1
0.34375
result
icons
339
model_pixel_error
0.07241242935923602, 0.06778112603307046, 0.07715227231417879
model_pixel_error_median
0.06746121495962143
exact_program_rate
0.02064896755162242
rendered_rate
1.0
blank_pixel_error
0.1720480552069557
selected_step
46000
nearest_training_icon_pixel_error
0.08973124639576185, 0.0848205749792233, 0.09466393787711656
model_minus_baseline_error_reduction
0.017318817036525845, 0.013095577734002774, 0.021401766072182183
icons_model_beats_baseline
234
clip_retrieval
model_greedy
top1_rate
0.5625
top5_rate
0.84375
mean_rank
3.375
clip_to_reference_mean
0.9102537911385298
chance_top1
0.03125
icons
32
nearest_training_icon
top1_rate
0.34375
top5_rate
0.59375
mean_rank
7.4375
clip_to_reference_mean
0.8629764690995216
chance_top1
0.03125
icons
32
gate_n_omnisvg_zero_shot_top1
0.609375
gate_n_omnisvg_best_of_12_top1
0.828125
model_calls_per_icon
692
predeclared_outcome
role
paired parent of the MUSICA scaling sweep (v10 arms); a rerun of the unchanged v9 config on an H100
against
r2s-full-v9-colour-621bc8e-b26ad95e-47646604
paired
icons
339
candidate_pixel_error
0.07241242935923602, 0.06769235660214266, 0.07717277473913198
parent_pixel_error
0.07194640226856078, 0.06741836603124769, 0.07659562947332754
parent_minus_candidate
-0.00046602709067524047, -0.003993854985468619, 0.0031151487177834727
candidate_better_icons
171
parent_better_icons
158
interval_excludes_zero_in_favour
false
reading
reproduces v9 within noise: the paired interval includes zero