masked-continuity-l13-start-head-full-f6befd5-2681icons-9b9b1699
completed —
Measured behaviour
modelposition-marginal floor
Table view
| optimizer step | model | position-marginal floor |
|---|---|---|
| 0 | 6.3217 | 4.7865 |
| 300 | 4.3265 | 4.7865 |
| 600 | 4.1690 | 4.7865 |
| 900 | 4.1284 | 4.7865 |
| 1200 | 4.0742 | 4.7865 |
| 1500 | 4.0403 | 4.7865 |
| 1800 | 4.0223 | 4.7865 |
| 2100 | 4.0048 | 4.7865 |
| 2400 | 3.9778 | 4.7865 |
| 2700 | 3.9505 | 4.7865 |
| 3000 | 3.8439 | 4.7865 |
| 3300 | 3.7794 | 4.7865 |
| 3600 | 3.7479 | 4.7865 |
| 3900 | 3.6923 | 4.7865 |
| 4200 | 3.6135 | 4.7865 |
| 4500 | 3.4984 | 4.7865 |
| 4800 | 3.4373 | 4.7865 |
| 5100 | 3.4076 | 4.7865 |
| 5400 | 3.3828 | 4.7865 |
| 5700 | 3.3552 | 4.7865 |
| 6000 | 3.3386 | 4.7865 |
| 6300 | 3.3086 | 4.7865 |
modelposition-marginal floor
Table view
| optimizer step | model | position-marginal floor |
|---|---|---|
| 0 | 0.0457 | 0.1307 |
| 300 | 0.1130 | 0.1307 |
| 600 | 0.1372 | 0.1307 |
| 900 | 0.1382 | 0.1307 |
| 1200 | 0.1517 | 0.1307 |
| 1500 | 0.1560 | 0.1307 |
| 1800 | 0.1668 | 0.1307 |
| 2100 | 0.1614 | 0.1307 |
| 2400 | 0.1646 | 0.1307 |
| 2700 | 0.1700 | 0.1307 |
| 3000 | 0.1684 | 0.1307 |
| 3300 | 0.1754 | 0.1307 |
| 3600 | 0.1802 | 0.1307 |
| 3900 | 0.1850 | 0.1307 |
| 4200 | 0.1813 | 0.1307 |
| 4500 | 0.1980 | 0.1307 |
| 4800 | 0.1845 | 0.1307 |
| 5100 | 0.1985 | 0.1307 |
| 5400 | 0.2006 | 0.1307 |
| 5700 | 0.2001 | 0.1307 |
| 6000 | 0.1910 | 0.1307 |
| 6300 | 0.2044 | 0.1307 |
Measured result
Verbatim from the run's summary.json.
checkpoint_round_trip
true
checkpoint_sha256
3b9e15b39acbe4fe…checks
all_valid
true
beats_copy_previous
true
config_sha256
be6adbabaa47b503…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
17.86283185840708
median_bins_off
10.5
n
339
coordinate_tau
1.0
criteria
beats_copy_previous
true
require_all_valid
true
deterministic_algorithms
true
device
cuda
evaluation_icons
339
evaluation_split
validation
first_train_loss
6.082370281219482
gpu
NVIDIA GeForce RTX 4080
held_out_nll_per_masked_token
3.3086356480574337
icon_presentations
100800
initial_held_out_nll
6.321668299633434
inpainting
all_valid
true
attempted
16
close_reproduction_rate
0.0
decoding
chain_order
false
greedy_for_criteria
true
iterations
8
samples
1
drop
median_rgba_mae
0.009685249485596707
exact_reproduction_rate
0.0
icons
16
marginal
helped
2
mean_recovery
-29608927.159026496
median_recovery
-1.3892468653280403
median_rgba_mae
0.026510208635923502
model
helped
1
mean_recovery
-0.7507010071727791
median_recovery
-0.20457199417362984
median_rgba_mae
0.014014085572500605
paired
drop_minus_model
interval
-0.009051582744172743, 0.0005392011325774868
interval_excludes_zero_above
false
mean
-0.004256190805797628
n
16
positive
1
marginal_minus_model
interval
-0.01302153221682321, 0.018151452143898985
interval_excludes_zero_above
false
mean
0.0025649599635378883
n
16
positive
13
rows_sha256
6a78e24b4fe2e76e…sheet_sha256
6d9aba7b0cda1497…last_train_loss
3.230076551437378
loss_reduction_factor
1.9106571324482382
marginal_masked_token_accuracy
0.1307154384077461
marginal_nll_per_masked_token
4.786513732286344
mask_mixture
drawn
span
100800
max_paths
2
random_rate
0.05, 0.95
span_max
1
weights
span
1.0
masked_token_accuracy
0.20441097364174288
metric_head
true
metrics_sha256
46bb6f6b39572424…model_parameters
531892
nll_ratio_to_marginal
0.6912412317423811
path_binding
false
predeclared_outcome
passed
schema_version
1
selected_step
6300
sequence_length
1376
start_features
true
steps_run
6300
study_version
masked-continuity-l13-start-head-full
timing
peak_vram_gib
1.9460439682006836
prepare_seconds
101.41797300000326
train_seconds
418.78929798598983
torch_version
2.14.0a0+4fdf77b940.nv26.08
train_icons
2681
vocabulary
418
Visual output
State transitions
- planned2026-09-21T09:39:00Z
- completed2026-09-21T09:57:15Zstart features and the metric head at the full budget beat the copy-the-previous-endpoint policy: 17.9 bins (median 10.5) against 25.6 (median 17.0) over one hidden segment in each of 339 held-out icons; held-out likelihood 3.309, 0.691 of the floor, still falling at step 6,300; 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-continuity-l13-start-head-full-f6befd5-2681icons-9b9b1699
config
configs/learning/masked-continuity-l13-start-head-full.yaml
outputs
checkpoint_root
data/processed/masked-continuity-l13-start-head-full
report_root
reports/learning/masked-continuity-l13-start-head-full