r2s-cur-rl-ctrl-s4-1df927d-be577bd0-47646604
completed —
Hypothesis
Supervised (control) fine-tuning of the final model for 800 steps at lr 2e-5 on the d02 recipe's online stream. Declared in state/CURRENT.md (2026-10-31 18:45): the RL arm's greedy float32 pixel error on the 339 validation renders beats the control's, seed-matched over seeds 7001-7004 with a bootstrap interval excluding zero, and the best RL run is more often recognisable than the final model on the same blind visual review.
Visual output
Written result
Hypothesis. Supervised (control) fine-tuning of the final model for 800 steps at lr 2e-5 on the d02 recipe's online stream. Declared in state/CURRENT.md (2026-10-31 18:45): the RL arm's greedy float32 pixel error on the 339 validation renders beats the control's, seed-matched over seeds 7001-7004 with a bootstrap interval excluding zero, and the best RL run is more often recognisable than the final model on the same blind visual review.
Result
| measure | value |
|---|---|
| icons evaluated | 339 (primary/validation) |
| model pixel error, mean [95% CI] | 0.0659 [0.0609, 0.0711] |
| model pixel error, median | 0.0584 |
| 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 | 0.0239 [0.0197, 0.0282] |
| icons where model beats nearest icon | 259 |
| CLIP top-1, model greedy (32 Gate N icons) | 0.781 |
| 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 | 191 |
| peak VRAM GiB | 2.303131580352783 |
| latency ms/icon, batch 1, median | 2149.9 |
| latency ms/icon, batch 1, p95 | 2168.6 |
| throughput ms/icon, batch 64 | 62.1 |
| model calls per icon (decoding) | 753.0 |
| 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
- running2026-10-31T17:53:05Z
- completed2026-10-31T17:58:23Z
Run record
Verbatim from runs/r2s-cur-rl-ctrl-s4-1df927d-be577bd0-47646604/run.yaml, the record committed before launch.
schema_version
2
run_id
r2s-cur-rl-ctrl-s4-1df927d-be577bd0-47646604
state
completed
planned_at
2026-10-31T17:53:05Z
hypothesis
Supervised (control) fine-tuning of the final model for 800 steps at lr 2e-5 on the d02 recipe's online stream. Declared in state/CURRENT.md (2026-10-31 18:45): the RL arm's greedy float32 pixel error on the 339 validation renders beats the control's, seed-matched over seeds 7001-7004 with a bootstrap interval excluding zero, and the best RL run is more often recognisable than the final model on the same blind visual review.
parent_run
r2s-cur-d02-s120k-seed2-53b3108-46294452-47646604
git_commit
1df927d22972b4037240b9f44ae2a00a55aad475
dirty_patch_sha256
config
configs/render2svg/cur-rl-ctrl-s4.yaml
config_sha256
be577bd0db48d25d…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
800
batch_size
16
learning_rate
2e-05
weight_decay
0.01
warmup_steps
50
seed
7004
eval_every
50
eval_icons
64
train_icons
None
evaluate_on_train
false
augment_variants
0
augment_seed
9004
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
degrade_probability
0.0
degrade_kinds
lineart, sloppy, flat
degrade_eval_seed
0
init_run
r2s-cur-d02-s120k-seed2-53b3108-46294452-47646604
init_sha256
38daeca4dd27fb4c…rl_samples
0
rl_temperature
1.0
rl_anchor
0.5
max_seconds
6000
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
evaluation_inputs
clean renders
init
run
r2s-cur-d02-s120k-seed2-53b3108-46294452-47646604
checkpoint
/home/eai-jdkxz/eai-team11/mojidiff/cache/runs/r2s-cur-d02-s120k-seed2-53b3108-46294452-47646604/best.pt
sha256
38daeca4dd27fb4c…step
64000
seed
7004
determinism
seeded; cuDNN and SDPA kernels not forced deterministic
model_parameters
9025826
command
.venv/bin/python -m mojidiff.learning.render2svg --config configs/render2svg/cur-rl-ctrl-s4.yaml
outputs
run_dir
runs/r2s-cur-rl-ctrl-s4-1df927d-be577bd0-47646604
checkpoints
/home/eai-jdkxz/eai-team11/mojidiff/cache/runs/r2s-cur-rl-ctrl-s4-1df927d-be577bd0-47646604
notes
stopped_at_step
800
completed_at
2026-10-31T17:58:23Z
resource
device
NVIDIA B300 SXM6 AC
torch_version
2.14.1+cu130
cuda_version
13.0
python_version
3.12.3
peak_vram_gib
2.303131580352783
train_seconds
190.8454976180219
inference_ms_per_icon
2149.926260986831
inference_p95_ms_per_icon
2168.6200829572044
icons_per_second
0.4651322318101241
batched_ms_per_icon
62.080325421447924
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
0.06585467074684513, 0.06086775280744239, 0.07110935816666916
model_pixel_error_median
0.058395616710186005
exact_program_rate
0.029498525073746312
rendered_rate
1.0
blank_pixel_error
0.1720480552069557
selected_step
500
nearest_training_icon_pixel_error
0.08973124639576185, 0.0848205749792233, 0.09466393787711656
model_minus_baseline_error_reduction
0.02387657564891672, 0.01969546831938288, 0.028184498844530833
icons_model_beats_baseline
259
clip_retrieval
model_greedy
top1_rate
0.78125
top5_rate
0.96875
mean_rank
1.71875
clip_to_reference_mean
0.9377066642045975
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
753