omnisvg-n1-image-finetune-self-207d732-32icons-9b9b1699
completed —
Measured behaviour
Table view
| optimizer step | model |
|---|---|
| 0 | 0.3581 |
| 50 | 0.2389 |
| 100 | 0.2261 |
| 150 | 0.2506 |
| 200 | 0.3111 |
| 250 | 0.3692 |
| 300 | 0.3906 |
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
0.24804076505824924
median
0.24212153255939484
n
64
clip_to_reference
mean
0.9180985242128372
median
0.9489547908306122
n
64
codec_valid_rate
0.375
config_sha256
89a3e6319948855f…control
clip_to_reference_mean
0.9180985242128372
control_clip_to_reference_mean
0.9046997493132949
control_codec_valid_rate
0.140625
control_reference_top1_rate
0.609375
control_rows_sha256
5cabd027e8c4e617…gain_ci95
-0.008400542545132337, 0.03505141753703355
icons_improved
17
mean_gain
0.013398774899542332
paired_icons
32
report_root
reports/learning/omnisvg-n1-image-control
criteria
clip_gain_over_control_ci_excludes_zero
true
max_memorised_exactly
0
min_ended_rate
0.9
min_reference_top1_rate
0.75
decoded_rate
1.0
drawings
64
ended_rate
0.6875
failures
encode:endpoint coordinate is outside 0..72 and clamping is disabled
2
encode:program needs 1 contours and up to 510 segments
2
encode:program needs 1 contours and up to 511 segments
1
encode:program needs 2 contours and up to 303 segments
1
encode:program needs 2 contours and up to 387 segments
1
encode:program needs 2 contours and up to 493 segments
1
encode:program needs 2 contours and up to 950 segments
1
encode:program needs 3 contours and up to 438 segments
1
encode:program needs 3 contours and up to 469 segments
1
encode:program needs 3 contours and up to 477 segments
1
encode:program needs 39 contours and up to 17 segments
1
encode:program needs 4 contours and up to 978 segments
1
encode:program needs 45 contours and up to 17 segments
1
encode:program needs 5 contours and up to 476 segments
1
encode:program needs 6 contours and up to 436 segments
1
encode:program needs 8 contours and up to 436 segments
1
encode:program needs 8 contours and up to 446 segments
1
encode:program needs 8 contours and up to 455 segments
1
encode:program needs 8 contours and up to 523 segments
1
encode:program needs 9 contours and up to 212 segments
1
pack:program needs 137 packed segments but capacity is 128
1
pack:program needs 146 packed segments but capacity is 128
1
pack:program needs 150 packed segments but capacity is 128
1
pack:program needs 155 packed segments but capacity is 128
1
pack:program needs 156 packed segments but capacity is 128
1
pack:program needs 159 packed segments but capacity is 128
2
pack:program needs 171 packed segments but capacity is 128
2
pack:program needs 183 packed segments but capacity is 128
1
pack:program needs 197 packed segments but capacity is 128
1
pack:program needs 200 packed segments but capacity is 128
1
pack:program needs 211 packed segments but capacity is 128
1
pack:program needs 215 packed segments but capacity is 128
2
pack:program needs 240 packed segments but capacity is 128
1
pack:program needs 281 packed segments but capacity is 128
1
pack:program needs 556 packed segments but capacity is 128
1
icons
32
max_new_tokens
2048
mean_snap_distance_rgb
5.869173621557129
median_paths
4.0
median_tokens
692.0
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
image
reference_rank
chance_top1
0.03125
mean
5.015625
scored
64
top1_rate
0.546875
top5_rate
0.78125
rerank
0
rows_sha256
cc8f44cf38e03aec…samples_per_icon
2
sampling
repetition_penalty
1.05
temperature
0.3
top_k
50
top_p
0.9
schema_version
1
sheet_sha256
1cfb197856e63137…study_version
omnisvg-n1-image-finetune-self
timing
generate_seconds
1277.0953915619175
peak_vram_gib
9.89112901687622
seconds_per_drawing
19.95461549315496
training
adapter_bytes
119816085
adapter_sha256
687ae75099968c21…condition
image
excluded_selection_over_max_tokens
0
excluded_train_over_max_tokens
0
held_out_nll
0.226131928882507
initial_held_out_nll
0.3580829583969109
load_seconds
37.41578564699739
max_tokens
2400
peak_vram_gib
9.89112901687622
selected_on
probe_clip_mean
selected_probe
probe_clip_mean
0.9124401956796646
probe_ended_rate
0.5
probe_median_tokens
976.5
selected_step
100
selection_encode_failures
17
selection_icons
15
sequences_per_step
8
split_max_dist
None
steps_run
300
token_source
self
train_encode_failures
358
train_icons
242
train_seconds
2024.3804181540036
Visual output
State transitions
- planned2026-09-22T20:49:26Z
- completed2026-09-23T04:50:31Ztrained on 242 of its own drawings (the 600 renders drawn greedily, 358 did not end and were not used), the probe stayed at 0.92, 0.95, 0.94, 0.94, 0.89, 0.91 with half the probe drawings ending - above every compact- and split-dialect probe - and the step-100 adapters were selected; on the 32 held-out icons: 69% end (control 50%), median 692 tokens (control 1,802), similarity 0.918 (control 0.905), paired gain +0.013 with an interval of -0.008 to +0.035, the right icon first for 35 of 64 (0.55; control 0.61); no training icon reproduced
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-n1-image-finetune-self-207d732-32icons-9b9b1699
config
configs/learning/omnisvg-n1-image-finetune-self.yaml
outputs
adapter_sha256
687ae75099968c21…report_root
reports/learning/omnisvg-n1-image-finetune-self