r2s-v10-s120k-5247da7-95eba274-47646604
completed —
Hypothesis
With training steps 60k -> 120k, and everything else as in v9, the render-to-program model transcribes the 339 validation renders with lower mean 72 px pixel error than v9 trained on the same machine and seed. Pass requires the paired 95% bootstrap interval on the per-icon reduction against that v9 rerun to exclude zero. Its greedy latency on the RTX 4080 is recorded; a pass that is too slow for the browser demo is a teacher for distillation, not the demo model.
Visual output
Written result
Hypothesis. With training steps 60k -> 120k, and everything else as in v9, the render-to-program model transcribes the 339 validation renders with lower mean 72 px pixel error than v9 trained on the same machine and seed. Pass requires the paired 95% bootstrap interval on the per-icon reduction against that v9 rerun to exclude zero. Its greedy latency on the RTX 4080 is recorded; a pass that is too slow for the browser demo is a teacher for distillation, not the demo model.
Result
| measure | value |
|---|---|
| icons evaluated | 339 (primary/validation) |
| model pixel error, mean [95% CI] | 0.0727 [0.0677, 0.0780] |
| model pixel error, median | 0.0678 |
| exact program rate | 0.032 |
| 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 | 0.0170 [0.0131, 0.0209] |
| icons where model beats nearest icon | 237 |
| CLIP top-1, model greedy (32 Gate N icons) | 0.750 |
| 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 | 6360 |
| peak VRAM GiB | 4.38838529586792 |
| latency ms/icon, batch 1, median | 2993.2 |
| latency ms/icon, batch 1, p95 | 8018.8 |
| throughput ms/icon, batch 64 | 79.3 |
| model calls per icon (decoding) | 763.0 |
| 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
- running2026-10-13T18:02:56Z
- completed2026-10-13T19:52:30Z
Run record
Verbatim from runs/r2s-v10-s120k-5247da7-95eba274-47646604/run.yaml, the record committed before launch.
schema_version
2
run_id
r2s-v10-s120k-5247da7-95eba274-47646604
state
completed
planned_at
2026-10-13T18:02:56Z
hypothesis
With training steps 60k -> 120k, and everything else as in v9, the render-to-program model transcribes the 339 validation renders with lower mean 72 px pixel error than v9 trained on the same machine and seed. Pass requires the paired 95% bootstrap interval on the per-icon reduction against that v9 rerun to exclude zero. Its greedy latency on the RTX 4080 is recorded; a pass that is too slow for the browser demo is a teacher for distillation, not the demo model.
parent_run
r2s-full-v9-colour
git_commit
5247da7ce9cb85d7952d2cf5833ef2daa3a416a1
dirty_patch_sha256
23cc26565a136aaa…config
configs/render2svg/v10-s120k.yaml
config_sha256
95eba27426e6e19a…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
120000
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/v10-s120k.yaml
outputs
run_dir
runs/r2s-v10-s120k-5247da7-95eba274-47646604
checkpoints
/data/fs201459/cp98556/predictable/cache/runs/r2s-v10-s120k-5247da7-95eba274-47646604
notes
completed_at
2026-10-13T19:52:30Z
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
6359.954050446046
inference_ms_per_icon
2993.2274209568277
inference_p95_ms_per_icon
8018.81466305349
icons_per_second
0.33408754476809377
batched_ms_per_icon
79.27600676521251
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
0.0727150620154145, 0.06768874085412906, 0.07802898193619236
model_pixel_error_median
0.067843496799469
exact_program_rate
0.032448377581120944
rendered_rate
1.0
blank_pixel_error
0.1720480552069557
selected_step
82000
nearest_training_icon_pixel_error
0.08973124639576185, 0.0848205749792233, 0.09466393787711656
model_minus_baseline_error_reduction
0.01701618438034736, 0.013114884193628393, 0.020865874362767867
icons_model_beats_baseline
237
clip_retrieval
model_greedy
top1_rate
0.75
top5_rate
0.9375
mean_rank
2.5625
clip_to_reference_mean
0.9251997601240873
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
763
predeclared_outcome
criterion
paired 95% bootstrap interval on the per-icon reduction against the v9 rerun excludes zero in favour
against
r2s-full-v9-colour-989aea8-b26ad95e-47646604
paired
icons
339
candidate_pixel_error
0.0727150620154145, 0.06770579691992493, 0.0781039449007262
parent_pixel_error
0.07241242935923602, 0.06769235660214266, 0.07717277473913198
parent_minus_candidate
-0.0003026326561784839, -0.003953152179811511, 0.0033242569278753445
candidate_better_icons
154
parent_better_icons
170
interval_excludes_zero_in_favour
false
passed
false
rtx4080_latency
not measured (no arm passed)