omnisvg-m1-finetune-52b1950-32icons-9b9b1699
completed —
Measured behaviour
Table view
| optimizer step | model |
|---|---|
| 0 | 2.9588 |
| 25 | 2.2915 |
| 50 | 1.6904 |
| 75 | 1.6012 |
| 100 | 1.5608 |
| 125 | 1.5264 |
| 150 | 1.4989 |
| 175 | 1.4856 |
| 200 | 1.4670 |
| 225 | 1.4485 |
| 250 | 1.4395 |
| 275 | 1.4415 |
| 300 | 1.4248 |
| 325 | 1.4132 |
| 350 | 1.4187 |
| 375 | 1.4109 |
| 400 | 1.4186 |
| 425 | 1.4244 |
| 450 | 1.3975 |
| 475 | 1.4170 |
| 500 | 1.4174 |
Measured result
Verbatim from the run's summary.json.
checks
clip_gain_over_control
false
ended_rate
false
memorisation
true
reference_top1_rate
false
clip
repo
openai/clip-vit-base-patch32
revision
3d74acf9a28c67741b2f4f2ea7635f0aaf6f0268
clip_to_caption
mean
0.22291783499531448
median
0.2209063321352005
n
64
clip_to_reference
mean
0.8184166345745325
median
0.8176865577697754
n
64
codec_valid_rate
0.828125
config_sha256
59fdc5a0962d8775…control
clip_to_reference_mean
0.8184166345745325
control_clip_to_reference_mean
0.8112438190728426
control_codec_valid_rate
0.734375
control_reference_top1_rate
0.078125
control_rows_sha256
3f6131b329d040a5…gain_ci95
-0.010761523968540132, 0.025219996599480505
icons_improved
15
mean_gain
0.007172815501689911
paired_icons
32
report_root
reports/learning/omnisvg-m1-control
criteria
clip_gain_over_control_ci_excludes_zero
true
max_memorised_exactly
0
min_ended_rate
0.9
min_reference_top1_rate
0.25
decoded_rate
1.0
drawings
64
ended_rate
0.890625
failures
encode:program needs 14 contours and up to 378 segments
1
encode:program needs 25 contours and up to 385 segments
1
encode:program needs 3 contours and up to 473 segments
1
encode:program needs 3 contours and up to 496 segments
1
encode:program needs 38 contours and up to 29 segments
1
encode:program needs 7 contours and up to 372 segments
1
encode:program needs 7 contours and up to 390 segments
1
encode:program needs 70 contours and up to 11 segments
1
encode:program needs 8 contours and up to 257 segments
1
pack:program needs 135 packed segments but capacity is 128
1
pack:program needs 189 packed segments but capacity is 128
1
icons
32
max_new_tokens
2048
mean_snap_distance_rgb
5.403812888905604
median_paths
2.0
median_tokens
120.0
memorised_exactly
0
model
checkpoint_sha256
39f36759fc677b32…fine_tuned
true
lora
alpha
32
dropout
0.05
rank
16
parameters
3876847616
repo
OmniSVG/OmniSVG1.1_4B
revision
117d4c4541839e02943554fe71c0a0d54fcef819
trainable_parameters
29933568
predeclared_outcome
falsified
prompt_style
training
reference_rank
chance_top1
0.03125
mean
11.546875
scored
64
top1_rate
0.0625
top5_rate
0.3125
rows_sha256
d62d7709d403ebe5…samples_per_icon
2
schema_version
1
sheet_sha256
9f622ab6571d37c2…study_version
omnisvg-m1-finetune
timing
generate_seconds
381.6565031869977
peak_vram_gib
9.167438983917236
seconds_per_drawing
5.9633828622968394
training
adapter_bytes
119816085
adapter_sha256
23bf01ba394a1e04…excluded_selection_over_max_tokens
0
excluded_train_over_max_tokens
0
held_out_nll
1.3975301341926876
initial_held_out_nll
2.9588475643566947
load_seconds
37.39460437602247
max_tokens
2048
peak_vram_gib
9.167438983917236
selected_step
450
selection_encode_failures
0
selection_icons
32
sequences_per_step
8
steps_run
500
train_encode_failures
0
train_icons
2681
train_seconds
1303.2937071160122
Visual output
State transitions
- planned2026-09-21T13:33:27Z
- completed2026-09-21T14:46:24Zheld-out likelihood falls from 2.96 to 1.40 nats a token (plateau from step 300 while train loss keeps falling to 0.85); the drawings: CLIP-to-reference 0.818 against the control's 0.811, paired gain +0.007 with a 95% interval of -0.011 to +0.025, 15 of 32 icons improved; the right icon ranked first for 4 of 64 (0.063, control 0.078), top five for 20; 89% ended; no training icon reproduced; median drawing 120 tokens against a training median of 647, with 16 drawings of exactly 62 tokens - the same blue square - and other repeated templates
Run record
This run has no run.yaml. What follows is the identity and configuration carried by its rows in state/runs.jsonl, the append-only registry.
run_id
omnisvg-m1-finetune-52b1950-32icons-9b9b1699
config
configs/learning/omnisvg-m1-finetune.yaml
outputs
adapter
data/processed/prior/omnisvg-m1-finetune/adapter.zip
adapter_sha256
23bf01ba394a1e04…report_root
reports/learning/omnisvg-m1-finetune