masked-inpaint-l14-start-head-ba54da5-2681icons-9b9b1699
completed —
Measured behaviour
modelposition-marginal floor
Table view
| optimizer step | model | position-marginal floor |
|---|---|---|
| 0 | 5.2606 | 4.1110 |
| 300 | 4.0376 | 4.1110 |
| 600 | 3.9234 | 4.1110 |
| 900 | 3.9053 | 4.1110 |
| 1200 | 3.9050 | 4.1110 |
| 1500 | 3.9001 | 4.1110 |
| 1800 | 3.9047 | 4.1110 |
| 2100 | 3.9040 | 4.1110 |
| 2400 | 3.8909 | 4.1110 |
| 2700 | 3.8980 | 4.1110 |
| 3000 | 3.8913 | 4.1110 |
| 3300 | 3.8865 | 4.1110 |
| 3600 | 3.8779 | 4.1110 |
| 3900 | 3.8962 | 4.1110 |
| 4200 | 3.8755 | 4.1110 |
| 4500 | 3.9024 | 4.1110 |
| 4800 | 3.8800 | 4.1110 |
| 5100 | 3.8800 | 4.1110 |
| 5400 | 3.9260 | 4.1110 |
| 5700 | 3.8952 | 4.1110 |
| 6000 | 3.8873 | 4.1110 |
| 6300 | 3.9047 | 4.1110 |
modelposition-marginal floor
Table view
| optimizer step | model | position-marginal floor |
|---|---|---|
| 0 | 0.0660 | 0.2413 |
| 300 | 0.2129 | 0.2413 |
| 600 | 0.2424 | 0.2413 |
| 900 | 0.2490 | 0.2413 |
| 1200 | 0.2495 | 0.2413 |
| 1500 | 0.2544 | 0.2413 |
| 1800 | 0.2538 | 0.2413 |
| 2100 | 0.2578 | 0.2413 |
| 2400 | 0.2568 | 0.2413 |
| 2700 | 0.2614 | 0.2413 |
| 3000 | 0.2605 | 0.2413 |
| 3300 | 0.2607 | 0.2413 |
| 3600 | 0.2630 | 0.2413 |
| 3900 | 0.2625 | 0.2413 |
| 4200 | 0.2617 | 0.2413 |
| 4500 | 0.2632 | 0.2413 |
| 4800 | 0.2667 | 0.2413 |
| 5100 | 0.2643 | 0.2413 |
| 5400 | 0.2637 | 0.2413 |
| 5700 | 0.2609 | 0.2413 |
| 6000 | 0.2632 | 0.2413 |
| 6300 | 0.2675 | 0.2413 |
Measured result
Verbatim from the run's summary.json.
checkpoint_round_trip
true
checkpoint_sha256
ec454fbde6509b8b…checks
all_valid
true
beats_drop_baseline
false
beats_marginal_baseline
true
median_recovery
false
config_sha256
0257980a69c181e9…continuity
copy_previous
mean_bins_off
25.64749262536873
median_bins_off
17.0
n
339
marginal
mean_bins_off
140.05162241887905
median_bins_off
138.5
n
339
model
mean_bins_off
41.79203539823009
median_bins_off
36.0
n
339
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
5.358754634857178
gpu
NVIDIA GeForce RTX 4080
held_out_nll_per_masked_token
3.875509736467909
icon_presentations
100800
initial_held_out_nll
5.260572421212839
inpainting
all_valid
true
attempted
64
close_reproduction_rate
0.0
decoding
chain_order
true
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
6
mean_recovery
-5055975.0345634995
median_recovery
-0.1087351943803935
median_rgba_mae
0.007291666666666667
paired
drop_minus_model
interval
-0.0016653814081545757, -0.0006072185317056271
interval_excludes_zero_above
false
mean
-0.0011362999699301014
n
64
positive
6
marginal_minus_model
interval
0.01057050893873887, 0.017620381253178944
interval_excludes_zero_above
true
mean
0.014095445095958907
n
64
positive
64
rows_sha256
a02b0457847a55fa…sheet_sha256
3b0f9e3f18bba2a7…last_train_loss
3.8391597270965576
loss_reduction_factor
1.3573885189119042
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
span_max
None
weights
geometry
0.1
path
0.3
random
0.3
span
0.2
style
0.1
masked_token_accuracy
0.2616700381532789
metric_head
true
metrics_sha256
fb1c4decc67f8d00…model_parameters
531892
nll_ratio_to_marginal
0.9427168654304083
path_binding
false
predeclared_outcome
falsified
schema_version
1
selected_step
4200
sequence_length
1376
start_features
true
steps_run
6300
study_version
masked-inpaint-l14-start-head
timing
peak_vram_gib
1.9460439682006836
prepare_seconds
103.3546569510072
train_seconds
421.51113984399126
torch_version
2.14.0a0+4fdf77b940.nv26.08
train_icons
2681
vocabulary
418
Visual output
State transitions
- planned2026-09-21T09:48:14Z
- completed2026-09-21T10:08:03Zwith start features, the metric head and chain-order decoding, whole-path completion still loses to the hole - mean paired difference -0.0011, interval [-0.0017, -0.0006], 6 of 64 helped, median recovery -0.11 - but the gap to the hole is a quarter of l2's -0.0040, the model beats the marginal policy on all 64 icons, and icon by icon it beats l2 on 44 of 64 by 0.0029; the fills are quiet rather than scribbles; 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-l14-start-head-ba54da5-2681icons-9b9b1699
config
configs/learning/masked-inpaint-l14-start-head.yaml
outputs
checkpoint_root
data/processed/masked-inpaint-l14-start-head
report_root
reports/learning/masked-inpaint-l14-start-head