omnisvg-n1-image-finetune-split-6ac69ad-32icons-9b9b1699
completed —
Measured behaviour
Table view
| optimizer step | model |
|---|---|
| 0 | 2.1793 |
| 50 | 1.4201 |
| 100 | 1.2602 |
| 150 | 1.1763 |
| 200 | 1.0497 |
| 250 | 1.0182 |
| 300 | 0.9882 |
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
0.2353658068459481
median
0.23205163329839706
n
64
clip_to_reference
mean
0.8742659306153655
median
0.8404034078121185
n
64
codec_valid_rate
0.421875
config_sha256
66808aeb87b190b0…control
clip_to_reference_mean
0.8742659306153655
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
-0.05153311442118138, -0.009521375549957157
icons_improved
8
mean_gain
-0.030433818697929382
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
0.359375
failures
encode:program needs 1 contours and up to 504 segments
1
encode:program needs 1 contours and up to 507 segments
1
encode:program needs 1 contours and up to 510 segments
7
encode:program needs 1 contours and up to 737 segments
1
encode:program needs 11 contours and up to 396 segments
1
encode:program needs 11 contours and up to 427 segments
1
encode:program needs 12 contours and up to 340 segments
1
encode:program needs 12 contours and up to 404 segments
1
encode:program needs 2 contours and up to 484 segments
1
encode:program needs 2 contours and up to 502 segments
1
encode:program needs 2 contours and up to 504 segments
1
encode:program needs 3 contours and up to 471 segments
1
encode:program needs 3 contours and up to 496 segments
2
encode:program needs 34 contours and up to 16 segments
1
encode:program needs 39 contours and up to 14 segments
2
encode:program needs 4 contours and up to 465 segments
1
encode:program needs 4 contours and up to 494 segments
1
encode:program needs 47 contours and up to 24 segments
1
encode:program needs 5 contours and up to 471 segments
1
encode:program needs 57 contours and up to 10 segments
1
encode:program needs 76 contours and up to 6 segments
1
encode:program needs 8 contours and up to 905 segments
1
encode:program needs 9 contours and up to 437 segments
1
normalize:geometry_error: M has sets of 2 args, 1 invalid
1
pack:program needs 176 packed segments but capacity is 128
1
pack:program needs 177 packed segments but capacity is 128
1
pack:program needs 239 packed segments but capacity is 128
1
pack:program needs 240 packed segments but capacity is 128
1
pack:program needs 316 packed segments but capacity is 128
1
icons
32
max_new_tokens
2048
mean_snap_distance_rgb
4.6385617039966185
median_paths
3.5
median_tokens
2048.0
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
8.53125
scored
64
top1_rate
0.46875
top5_rate
0.515625
rerank
0
rows_sha256
e61ac0b209cbb29e…samples_per_icon
2
sampling
repetition_penalty
1.0
temperature
0.3
top_k
50
top_p
0.9
schema_version
1
sheet_sha256
0386caf964a8d1a7…study_version
omnisvg-n1-image-finetune-split
timing
generate_seconds
1869.2730375102255
peak_vram_gib
9.898649215698242
seconds_per_drawing
29.207391211097274
training
adapter_bytes
119816085
adapter_sha256
5efb6566121bc3e9…condition
image
excluded_selection_over_max_tokens
0
excluded_train_over_max_tokens
187
held_out_nll
0.988194631040175
initial_held_out_nll
2.1792550907944137
load_seconds
37.49486518499907
max_tokens
2400
peak_vram_gib
9.898649215698242
selected_on
probe_clip_mean
selected_probe
probe_clip_mean
0.8453782126307487
probe_ended_rate
0.0
probe_median_tokens
1536.0
selected_step
0
selection_encode_failures
0
selection_icons
32
sequences_per_step
8
split_max_dist
10.0
steps_run
300
token_source
openmoji
train_encode_failures
0
train_icons
2494
train_seconds
3545.0822290300275
Visual output
State transitions
- planned2026-09-22T20:23:52Z
- completed2026-09-22T22:56:37Zin the released model's own dialect (segments of at most 10 units; 2,494 icons under the cap) the free-running probe again never beat zero-shot - 0.78, 0.78, 0.80, 0.80, 0.83, 0.85 at steps 50 to 300, no probe drawing ending - while held-out likelihood fell from 2.18 to 0.99; training stopped on patience at step 300 and the initial adapters were restored, so the evaluation reads the released model without its repetition penalty: 36% end, the right icon first for 30 of 64 (0.47; the control with the penalty 0.61), similarity 0.874, paired gain -0.030
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-split-6ac69ad-32icons-9b9b1699
config
configs/learning/omnisvg-n1-image-finetune-split.yaml
outputs
adapter_sha256
5efb6566121bc3e9…report_root
reports/learning/omnisvg-n1-image-finetune-split