omnisvg-n1-image-finetune-merger-ff32137-32icons-9b9b1699
completed —
Measured behaviour
Table view
| optimizer step | model |
|---|---|
| 0 | 2.5848 |
| 50 | 1.3847 |
| 100 | 1.2510 |
| 150 | 1.2092 |
| 200 | 1.1684 |
| 250 | 1.1500 |
| 300 | 1.1288 |
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
0.23122292128391564
median
0.2238304391503334
n
64
clip_to_reference
mean
0.8507985072210431
median
0.8308261036872864
n
64
codec_valid_rate
0.25
config_sha256
cebb8b4bf09ae513…control
clip_to_reference_mean
0.8507985072210431
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
-0.0766045211115852, -0.03149892485234887
icons_improved
6
mean_gain
-0.05390124209225178
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
0.140625
failures
encode:program needs 1 contours and up to 1022 segments
1
encode:program needs 1 contours and up to 511 segments
7
encode:program needs 1 contours and up to 512 segments
1
encode:program needs 1 contours and up to 515 segments
1
encode:program needs 1 contours and up to 779 segments
1
encode:program needs 102 contours and up to 8 segments
2
encode:program needs 112 contours and up to 9 segments
1
encode:program needs 114 contours and up to 9 segments
1
encode:program needs 12 contours and up to 292 segments
1
encode:program needs 127 contours and up to 9 segments
1
encode:program needs 128 contours and up to 7 segments
5
encode:program needs 20 contours and up to 307 segments
1
encode:program needs 3 contours and up to 615 segments
1
encode:program needs 4 contours and up to 461 segments
1
encode:program needs 4 contours and up to 475 segments
1
encode:program needs 40 contours and up to 15 segments
2
encode:program needs 43 contours and up to 14 segments
1
encode:program needs 47 contours and up to 13 segments
1
encode:program needs 51 contours and up to 45 segments
1
encode:program needs 57 contours and up to 11 segments
3
encode:program needs 6 contours and up to 974 segments
1
encode:program needs 84 contours and up to 11 segments
2
encode:program needs 86 contours and up to 7 segments
1
encode:program needs 94 contours and up to 41 segments
1
normalize:geometry_error: M has sets of 2 args, 1 invalid
4
pack:program needs 137 packed segments but capacity is 128
1
pack:program needs 230 packed segments but capacity is 128
1
pack:program needs 332 packed segments but capacity is 128
1
pack:program needs 539 packed segments but capacity is 128
2
icons
32
max_new_tokens
2048
mean_snap_distance_rgb
5.564501678527245
median_paths
29.0
median_tokens
2048.0
memorised_exactly
0
model
checkpoint_sha256
39f36759fc677b32…fine_tuned
true
lora
alpha
32
dropout
0.05
merger
true
rank
16
parameters
3877126144
repo
OmniSVG/OmniSVG1.1_4B
revision
117d4c4541839e02943554fe71c0a0d54fcef819
trainable_parameters
30212096
predeclared_outcome
falsified
prompt_style
image
reference_rank
chance_top1
0.03125
mean
9.40625
scored
64
top1_rate
0.21875
top5_rate
0.390625
rows_sha256
7f5f35f42c0368d0…samples_per_icon
2
sampling
repetition_penalty
1.0
temperature
0.3
top_k
50
top_p
0.9
schema_version
1
sheet_sha256
d2376f94aaa81a81…study_version
omnisvg-n1-image-finetune-merger
timing
generate_seconds
2268.6915475720307
peak_vram_gib
9.890161037445068
seconds_per_drawing
35.44830543081298
training
adapter_bytes
120930749
adapter_sha256
f4b7548ff9e15425…condition
image
excluded_selection_over_max_tokens
0
excluded_train_over_max_tokens
0
held_out_nll
1.1288179577751054
initial_held_out_nll
2.584805574869435
load_seconds
37.41137047400116
max_tokens
2400
peak_vram_gib
9.890161037445068
selected_on
probe_clip_mean
selected_probe
probe_clip_mean
0.8543731793761253
probe_ended_rate
0.125
probe_median_tokens
1024.0
selected_step
0
selection_encode_failures
0
selection_icons
32
sequences_per_step
8
steps_run
300
train_encode_failures
0
train_icons
2681
train_seconds
2326.9490019879886
Visual output
State transitions
- planned2026-09-22T18:59:38Z
- completed2026-09-22T20:23:11Zthe free-running probe (8 selection icons, greedy, 1,024 tokens) never beat the zero-shot model at any of six checkpoints - similarity 0.80 to 0.85 against the zero-shot probe's, with at most 1 of 8 drawings ending - and training stopped on patience at step 300; the step-300 adapters draw worse than zero-shot on the 32 held-out icons: 14% end, the right icon first for 14 of 64 (0.22; control 0.61), similarity 0.851, paired gain -0.054 with an interval of -0.077 to -0.032
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-merger-ff32137-32icons-9b9b1699
config
configs/learning/omnisvg-n1-image-finetune-merger.yaml
outputs
adapter_sha256
f4b7548ff9e15425…report_root
reports/learning/omnisvg-n1-image-finetune-merger