masked-inpaint-l4-binding-438a98f-2681icons-9b9b1699
completed —
Measured behaviour
modelposition-marginal floor
Table view
| optimizer step | model | position-marginal floor |
|---|---|---|
| 0 | 4.8657 | 4.1110 |
| 300 | 3.9562 | 4.1110 |
| 600 | 3.9254 | 4.1110 |
| 900 | 3.9005 | 4.1110 |
| 1200 | 3.8959 | 4.1110 |
| 1500 | 3.8872 | 4.1110 |
| 1800 | 3.8836 | 4.1110 |
| 2100 | 3.8756 | 4.1110 |
| 2400 | 3.8703 | 4.1110 |
| 2700 | 3.8645 | 4.1110 |
| 3000 | 3.8538 | 4.1110 |
| 3300 | 3.8726 | 4.1110 |
| 3600 | 3.8564 | 4.1110 |
| 3900 | 3.8566 | 4.1110 |
| 4200 | 3.8573 | 4.1110 |
| 4500 | 3.8771 | 4.1110 |
| 4800 | 3.8556 | 4.1110 |
| 5100 | 3.8513 | 4.1110 |
| 5400 | 3.8727 | 4.1110 |
| 5700 | 3.8727 | 4.1110 |
| 6000 | 3.8431 | 4.1110 |
| 6300 | 3.8542 | 4.1110 |
modelposition-marginal floor
Table view
| optimizer step | model | position-marginal floor |
|---|---|---|
| 0 | 0.0367 | 0.2413 |
| 300 | 0.2417 | 0.2413 |
| 600 | 0.2457 | 0.2413 |
| 900 | 0.2545 | 0.2413 |
| 1200 | 0.2522 | 0.2413 |
| 1500 | 0.2571 | 0.2413 |
| 1800 | 0.2591 | 0.2413 |
| 2100 | 0.2619 | 0.2413 |
| 2400 | 0.2606 | 0.2413 |
| 2700 | 0.2623 | 0.2413 |
| 3000 | 0.2652 | 0.2413 |
| 3300 | 0.2660 | 0.2413 |
| 3600 | 0.2676 | 0.2413 |
| 3900 | 0.2674 | 0.2413 |
| 4200 | 0.2691 | 0.2413 |
| 4500 | 0.2680 | 0.2413 |
| 4800 | 0.2715 | 0.2413 |
| 5100 | 0.2717 | 0.2413 |
| 5400 | 0.2744 | 0.2413 |
| 5700 | 0.2718 | 0.2413 |
| 6000 | 0.2739 | 0.2413 |
| 6300 | 0.2751 | 0.2413 |
Measured result
Verbatim from the run's summary.json.
checkpoint_round_trip
true
checkpoint_sha256
6a1f28fe5fe93501…checks
all_valid
true
beats_drop_baseline
false
beats_marginal_baseline
true
median_recovery
false
config_sha256
27f9a1a73dde47af…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
4.834274768829346
gpu
NVIDIA GeForce RTX 4080
held_out_nll_per_masked_token
3.8430836629468286
icon_presentations
100800
initial_held_out_nll
4.865704408062041
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
8
mean_recovery
-8971332.327989401
median_recovery
-0.8093105148673071
median_rgba_mae
0.013239454732510288
paired
drop_minus_model
interval
-0.006946371188264783, -0.0022920140771067977
interval_excludes_zero_above
false
mean
-0.00461919263268579
n
64
positive
8
marginal_minus_model
interval
0.006594306252434617, 0.014630798613971825
interval_excludes_zero_above
true
mean
0.01061255243320322
n
64
positive
51
rows_sha256
cbcff1b8671bd7ed…sheet_sha256
107003c67933b6c1…last_train_loss
3.869515895843506
loss_reduction_factor
1.2660938024781587
marginal_masked_token_accuracy
0.2413130873428418
marginal_nll_per_masked_token
4.1110007453812765
mask_mixture
drawn
geometry
10032
path
30461
random
30235
span
20052
style
10020
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
0.27388420863954116
metrics_sha256
d4ad63137fea5f15…model_parameters
542050
nll_ratio_to_marginal
0.9348292303922704
path_binding
true
predeclared_outcome
falsified
schema_version
1
selected_step
6000
sequence_length
1376
steps_run
6300
study_version
masked-inpaint-l4-binding
timing
peak_vram_gib
1.9463090896606445
prepare_seconds
101.93874733400298
train_seconds
395.4905390780186
torch_version
2.14.0a0+4fdf77b940.nv26.08
train_icons
2681
vocabulary
418
Visual output
State transitions
- planned2026-09-21T08:35:58Z
- completed2026-09-21T08:47:17Zindistinguishable from l2 on every measure: held-out likelihood 3.843 against 3.869, accuracy 0.274 against 0.266, 8 icons helped against 7, worse than leaving the hole on 56 of 64, and icon by icon the two arms differ by 0.0006 RGBA MAE with binding better on 25 of 64; 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-l4-binding-438a98f-2681icons-9b9b1699
config
configs/learning/masked-inpaint-l4-binding.yaml
outputs
checkpoint_root
data/processed/masked-inpaint-l4-binding
report_root
reports/learning/masked-inpaint-l4-binding