masked-inpaint-l2-54061b2-2681icons-9b9b1699
completed —
Measured behaviour
modelposition-marginal floor
Table view
| optimizer step | model | position-marginal floor |
|---|---|---|
| 0 | 4.8050 | 4.1110 |
| 300 | 3.9821 | 4.1110 |
| 600 | 3.9263 | 4.1110 |
| 900 | 3.9050 | 4.1110 |
| 1200 | 3.9054 | 4.1110 |
| 1500 | 3.8983 | 4.1110 |
| 1800 | 3.8985 | 4.1110 |
| 2100 | 3.8932 | 4.1110 |
| 2400 | 3.8982 | 4.1110 |
| 2700 | 3.8869 | 4.1110 |
| 3000 | 3.8989 | 4.1110 |
| 3300 | 3.8688 | 4.1110 |
| 3600 | 3.8978 | 4.1110 |
| 3900 | 3.9096 | 4.1110 |
| 4200 | 3.9160 | 4.1110 |
| 4500 | 3.9050 | 4.1110 |
| 4800 | 3.8914 | 4.1110 |
| 5100 | 3.8874 | 4.1110 |
| 5400 | 3.9847 | 4.1110 |
| 5700 | 3.8814 | 4.1110 |
modelposition-marginal floor
Table view
| optimizer step | model | position-marginal floor |
|---|---|---|
| 0 | 0.0490 | 0.2413 |
| 300 | 0.2387 | 0.2413 |
| 600 | 0.2463 | 0.2413 |
| 900 | 0.2525 | 0.2413 |
| 1200 | 0.2505 | 0.2413 |
| 1500 | 0.2545 | 0.2413 |
| 1800 | 0.2562 | 0.2413 |
| 2100 | 0.2596 | 0.2413 |
| 2400 | 0.2586 | 0.2413 |
| 2700 | 0.2629 | 0.2413 |
| 3000 | 0.2617 | 0.2413 |
| 3300 | 0.2661 | 0.2413 |
| 3600 | 0.2637 | 0.2413 |
| 3900 | 0.2643 | 0.2413 |
| 4200 | 0.2657 | 0.2413 |
| 4500 | 0.2670 | 0.2413 |
| 4800 | 0.2701 | 0.2413 |
| 5100 | 0.2689 | 0.2413 |
| 5400 | 0.2707 | 0.2413 |
| 5700 | 0.2701 | 0.2413 |
Measured result
Verbatim from the run's summary.json.
checkpoint_round_trip
true
checkpoint_sha256
45d32c26f2dba0af…checks
all_valid
true
beats_drop_baseline
false
beats_marginal_baseline
true
median_recovery
false
config_sha256
2a8a71d97359a9d9…coordinate_tau
1.0
criteria
beats_drop_baseline
true
beats_marginal_baseline
true
min_median_recovery
0.3
require_all_valid
true
deterministic_algorithms
true
device
cuda
evaluation_icons
339
evaluation_split
validation
first_train_loss
4.784581661224365
gpu
NVIDIA GeForce RTX 4080
held_out_nll_per_masked_token
3.8688370011586817
icon_presentations
91200
initial_held_out_nll
4.805001695162264
inpainting
all_valid
true
attempted
64
decoding
greedy_for_criteria
true
iterations
8
samples
1
drop
median_rgba_mae
0.006453019093439845
exact_reproduction_rate
0.0
icons
64
marginal
helped
0
mean_recovery
-112236156.97001305
median_recovery
-1.7608338149263418
median_rgba_mae
0.019917714385136773
model
helped
7
mean_recovery
-165497.3872923598
median_recovery
-0.6925183889869049
median_rgba_mae
0.01281734144274994
paired
drop_minus_model
interval
-0.006853281016532452, -0.0012139148046972728
interval_excludes_zero_above
false
mean
-0.004033597910614862
n
64
positive
7
marginal_minus_model
interval
0.0065920744050887315, 0.015804219905459563
interval_excludes_zero_above
true
mean
0.011198147155274146
n
64
positive
54
rows_sha256
7924e3ba6d562112…sheet_sha256
521d238cb1d1d221…last_train_loss
3.75107479095459
loss_reduction_factor
1.2419757394077882
marginal_masked_token_accuracy
0.2413130873428418
marginal_nll_per_masked_token
4.1110007453812765
mask_mixture
drawn
geometry
9084
path
27531
random
27359
span
18210
style
9016
max_paths
2
random_rate
0.05, 0.95
weights
geometry
0.1
path
0.3
random
0.3
span
0.2
style
0.1
masked_token_accuracy
0.266074309272015
metrics_sha256
cbbad9bb94a87b3e…model_parameters
526498
nll_ratio_to_marginal
0.941093724078093
predeclared_outcome
falsified
schema_version
1
selected_step
3300
sequence_length
1376
steps_run
5700
study_version
masked-inpaint-l2
timing
peak_vram_gib
1.9457483291625977
prepare_seconds
101.68371957101044
train_seconds
345.6692114850157
torch_version
2.14.0a0+4fdf77b940.nv26.08
train_icons
2681
vocabulary
418
Visual output
State transitions
- planned2026-09-21T08:16:40Z
- completed2026-09-21T08:30:48Zthe model beats the position-marginal policy on 54 of 64 held-out icons with the interval excluding zero, and is worse than leaving the hole on 57 of 64: mean paired difference -0.0040 RGBA MAE, interval [-0.0069, -0.0012]; median recovery -0.69 against a bar of 0.30; every completion valid
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
masked-inpaint-l2-54061b2-2681icons-9b9b1699
config
configs/learning/masked-inpaint-l2.yaml
outputs
checkpoint_root
data/processed/masked-inpaint-l2
report_root
reports/learning/masked-inpaint-l2