r2s-cur-d02-s240k-53b3108-52ce75ff-47646604
completed —
Hypothesis
With dropout 0.2, 240k steps transcribe validation renders better than v9. Pass requires the paired 95% bootstrap interval on the per-icon reduction in mean 72 px pixel error on the 339 validation renders to exclude zero against each normal v9 run (RTX 4080 original, H100 rerun, B300 seeds 2 and 3). Its greedy latency on the RTX 4080 is recorded.
Visual output
Written result
Hypothesis. With dropout 0.2, 240k steps transcribe validation renders better than v9. Pass requires the paired 95% bootstrap interval on the per-icon reduction in mean 72 px pixel error on the 339 validation renders to exclude zero against each normal v9 run (RTX 4080 original, H100 rerun, B300 seeds 2 and 3). Its greedy latency on the RTX 4080 is recorded.
Result
| measure | value |
|---|---|
| icons evaluated | 339 (primary/validation) |
| model pixel error, mean [95% CI] | 0.0700 [0.0648, 0.0754] |
| model pixel error, median | 0.0595 |
| exact program rate | 0.029 |
| 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 | 0.0198 [0.0152, 0.0243] |
| icons where model beats nearest icon | 241 |
| 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 B300 SXM6 AC |
| parameters | 9,025,826 |
| train seconds | 10450 |
| peak VRAM GiB | 4.388660430908203 |
| latency ms/icon, batch 1, median | 2927.1 |
| latency ms/icon, batch 1, p95 | 2935.1 |
| throughput ms/icon, batch 64 | 95.5 |
| model calls per icon (decoding) | 587.0 |
| 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
- running2026-10-30T18:18:03Z
- completed2026-10-30T21:14:32Z
Run record
Verbatim from runs/r2s-cur-d02-s240k-53b3108-52ce75ff-47646604/run.yaml, the record committed before launch.
schema_version
2
run_id
r2s-cur-d02-s240k-53b3108-52ce75ff-47646604
state
completed
planned_at
2026-10-30T18:18:03Z
hypothesis
With dropout 0.2, 240k steps transcribe validation renders better than v9. Pass requires the paired 95% bootstrap interval on the per-icon reduction in mean 72 px pixel error on the 339 validation renders to exclude zero against each normal v9 run (RTX 4080 original, H100 rerun, B300 seeds 2 and 3). Its greedy latency on the RTX 4080 is recorded.
parent_run
r2s-cur-d02-s120k
git_commit
53b3108784dee11612fef24a88aa39bd393843c8
dirty_patch_sha256
config
configs/render2svg/cur-d02-s240k.yaml
config_sha256
52ce75ff3ba84fb1…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
240000
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/cur-d02-s240k.yaml
outputs
run_dir
runs/r2s-cur-d02-s240k-53b3108-52ce75ff-47646604
checkpoints
/home/eai-jdkxz/eai-team11/mojidiff/cache/runs/r2s-cur-d02-s240k-53b3108-52ce75ff-47646604
notes
completed_at
2026-10-30T21:14:32Z
resource
device
NVIDIA B300 SXM6 AC
torch_version
2.14.1+cu130
cuda_version
13.0
python_version
3.12.3
peak_vram_gib
4.388660430908203
train_seconds
10449.59109789203
inference_ms_per_icon
2927.0850614993833
inference_p95_ms_per_icon
2935.1247169543058
icons_per_second
0.3416368089719113
batched_ms_per_icon
95.45068140687363
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
0.06995163615836344, 0.06478176118722481, 0.07542944916446329
model_pixel_error_median
0.0595308355987072
exact_program_rate
0.029498525073746312
rendered_rate
1.0
blank_pixel_error
0.1720480552069557
selected_step
56000
nearest_training_icon_pixel_error
0.08973124639576185, 0.0848205749792233, 0.09466393787711656
model_minus_baseline_error_reduction
0.019779610237398406, 0.01522319130013076, 0.02434192330937412
icons_model_beats_baseline
241
clip_retrieval
model_greedy
top1_rate
0.5625
top5_rate
0.96875
mean_rank
2.5
clip_to_reference_mean
0.9225332550704479
chance_top1
0.03125
icons
32
nearest_training_icon
top1_rate
0.34375
top5_rate
0.59375
mean_rank
7.40625
clip_to_reference_mean
0.8630617223680019
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
587