r2s-overfit4-940f5d3-f7306d2b-47646604
completed —
Hypothesis
An 8.9M-parameter render-to-program encoder-decoder, trained on four OpenMoji icons, decodes those four icons greedily from their 144 px renders to programs whose 72 px renders are within 0.01 mean RGB error of the originals, in under 1,000 steps.
Visual output
Written result
Hypothesis. An 8.9M-parameter render-to-program encoder-decoder, trained on four OpenMoji icons, decodes those four icons greedily from their 144 px renders to programs whose 72 px renders are within 0.01 mean RGB error of the originals, in under 1,000 steps.
Result
| measure | value |
|---|---|
| icons evaluated | 4 (primary/train) |
| model pixel error, mean [95% CI] | 0.0000 [0.0000, 0.0000] |
| model pixel error, median | 0.0000 |
| exact program rate | 1.000 |
| rendered rate | 1.000 |
| blank canvas pixel error | 0.1657 |
Resources
| measure | value |
|---|---|
| device | NVIDIA GeForce RTX 4080 |
| parameters | 8,921,634 |
| train seconds | 14 |
| peak VRAM GiB | 0.6286492347717285 |
| latency ms/icon, batch 1, median | 147.9 |
| latency ms/icon, batch 1, p95 | 149.0 |
| throughput ms/icon, batch 4 | 56.0 |
| model calls per icon (decoding) | 97.0 |
| 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
- running2026-09-27T19:57:24Z
- completed2026-09-27T19:57:42Z
Run record
Verbatim from runs/r2s-overfit4-940f5d3-f7306d2b-47646604/run.yaml, the record committed before launch.
schema_version
2
run_id
r2s-overfit4-940f5d3-f7306d2b-47646604
state
completed
planned_at
2026-09-27T19:57:24Z
hypothesis
An 8.9M-parameter render-to-program encoder-decoder, trained on four OpenMoji icons, decodes those four icons greedily from their 144 px renders to programs whose 72 px renders are within 0.01 mean RGB error of the originals, in under 1,000 steps.
parent_run
None
git_commit
940f5d39cfd21468836064e15eca5de3f8831f73
dirty_patch_sha256
config
configs/render2svg/overfit-4.yaml
config_sha256
f7306d2b7a6e5039…config_resolved
model
image_size
144
d_model
256
heads
8
encoder_layers
2
decoder_layers
6
feedforward
1024
dropout
0.0
training
steps
800
batch_size
4
learning_rate
0.001
weight_decay
0.01
warmup_steps
50
seed
7001
eval_every
100
eval_icons
4
train_icons
4
evaluate_on_train
true
augment_mirror
0.0
bf16
true
final_eval_icons
4
clip_icons
0
dataset
pilot_config
configs/learning/openmoji-g1-geometric-gate-v16.yaml
cache_sha256
476466042da98d72…train_icons
4
evaluated_on
primary/train
seed
7001
determinism
seeded; cuDNN and SDPA kernels not forced deterministic
model_parameters
8921634
command
.venv/bin/python -m mojidiff.learning.render2svg --config configs/render2svg/overfit-4.yaml
outputs
run_dir
runs/r2s-overfit4-940f5d3-f7306d2b-47646604
checkpoints
/home/dev/.cache/mojidiff/runs/r2s-overfit4-940f5d3-f7306d2b-47646604
notes
completed_at
2026-09-27T19:57:42Z
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
0.6286492347717285
train_seconds
14.339369946043007
inference_ms_per_icon
147.90797646855935
inference_p95_ms_per_icon
148.98352301679552
icons_per_second
6.7609605910102415
batched_ms_per_icon
55.956157768378034
batched_icons_per_call
4
excludes
rendering the output SVG to pixels
baselines
blank_canvas_pixel_error
0.16571136377751827
result
icons
4
model_pixel_error
0.0, 0.0, 0.0
model_pixel_error_median
0.0
exact_program_rate
1.0
rendered_rate
1.0
blank_pixel_error
0.16571136377751827
selected_step
100
model_calls_per_icon
97