masked-span-l21-specialist-4-long-5a50b51-2681icons-9b9b1699
completed —
Measured behaviour
modelposition-marginal floor
Table view
| optimizer step | model | position-marginal floor |
|---|---|---|
| 0 | 6.2190 | 4.7651 |
| 300 | 4.4374 | 4.7651 |
| 600 | 4.3458 | 4.7651 |
| 900 | 4.2691 | 4.7651 |
| 1200 | 4.2303 | 4.7651 |
| 1500 | 4.2036 | 4.7651 |
| 1800 | 4.1980 | 4.7651 |
| 2100 | 4.1683 | 4.7651 |
| 2400 | 4.1551 | 4.7651 |
| 2700 | 4.1356 | 4.7651 |
| 3000 | 4.1411 | 4.7651 |
| 3300 | 4.1260 | 4.7651 |
| 3600 | 4.1054 | 4.7651 |
| 3900 | 4.0552 | 4.7651 |
| 4200 | 4.0108 | 4.7651 |
| 4500 | 3.9464 | 4.7651 |
| 4800 | 3.9314 | 4.7651 |
| 5100 | 3.8987 | 4.7651 |
| 5400 | 3.9052 | 4.7651 |
| 5700 | 3.8996 | 4.7651 |
| 6000 | 3.8908 | 4.7651 |
| 6300 | 3.8552 | 4.7651 |
| 6600 | 3.8397 | 4.7651 |
| 6900 | 3.8406 | 4.7651 |
| 7200 | 3.8023 | 4.7651 |
| 7500 | 3.7994 | 4.7651 |
| 7800 | 3.7683 | 4.7651 |
| 8100 | 3.7637 | 4.7651 |
| 8400 | 3.7838 | 4.7651 |
| 8700 | 3.7607 | 4.7651 |
| 9000 | 3.7278 | 4.7651 |
| 9300 | 3.7337 | 4.7651 |
| 9600 | 3.7029 | 4.7651 |
| 9900 | 3.7325 | 4.7651 |
| 10200 | 3.7127 | 4.7651 |
| 10500 | 3.6997 | 4.7651 |
| 10800 | 3.6716 | 4.7651 |
| 11100 | 3.7083 | 4.7651 |
| 11400 | 3.6752 | 4.7651 |
| 11700 | 3.6990 | 4.7651 |
| 12000 | 3.6701 | 4.7651 |
| 12300 | 3.6453 | 4.7651 |
| 12600 | 3.7005 | 4.7651 |
| 12900 | 3.6662 | 4.7651 |
| 13200 | 3.6500 | 4.7651 |
| 13500 | 3.6806 | 4.7651 |
| 13800 | 3.6443 | 4.7651 |
| 14100 | 3.6249 | 4.7651 |
| 14400 | 3.6228 | 4.7651 |
| 14700 | 3.6288 | 4.7651 |
| 15000 | 3.6066 | 4.7651 |
| 15300 | 3.6106 | 4.7651 |
| 15600 | 3.5931 | 4.7651 |
| 15900 | 3.5950 | 4.7651 |
| 16200 | 3.6270 | 4.7651 |
| 16500 | 3.6099 | 4.7651 |
| 16800 | 3.6189 | 4.7651 |
| 17100 | 3.6168 | 4.7651 |
| 17400 | 3.6098 | 4.7651 |
| 17700 | 3.6195 | 4.7651 |
| 18000 | 3.5987 | 4.7651 |
modelposition-marginal floor
Table view
| optimizer step | model | position-marginal floor |
|---|---|---|
| 0 | 0.0520 | 0.1280 |
| 300 | 0.1386 | 0.1280 |
| 600 | 0.1406 | 0.1280 |
| 900 | 0.1467 | 0.1280 |
| 1200 | 0.1532 | 0.1280 |
| 1500 | 0.1608 | 0.1280 |
| 1800 | 0.1511 | 0.1280 |
| 2100 | 0.1602 | 0.1280 |
| 2400 | 0.1587 | 0.1280 |
| 2700 | 0.1646 | 0.1280 |
| 3000 | 0.1590 | 0.1280 |
| 3300 | 0.1614 | 0.1280 |
| 3600 | 0.1631 | 0.1280 |
| 3900 | 0.1611 | 0.1280 |
| 4200 | 0.1631 | 0.1280 |
| 4500 | 0.1789 | 0.1280 |
| 4800 | 0.1628 | 0.1280 |
| 5100 | 0.1663 | 0.1280 |
| 5400 | 0.1786 | 0.1280 |
| 5700 | 0.1730 | 0.1280 |
| 6000 | 0.1713 | 0.1280 |
| 6300 | 0.1774 | 0.1280 |
| 6600 | 0.1719 | 0.1280 |
| 6900 | 0.1690 | 0.1280 |
| 7200 | 0.1821 | 0.1280 |
| 7500 | 0.1783 | 0.1280 |
| 7800 | 0.1710 | 0.1280 |
| 8100 | 0.1789 | 0.1280 |
| 8400 | 0.1906 | 0.1280 |
| 8700 | 0.1824 | 0.1280 |
| 9000 | 0.1882 | 0.1280 |
| 9300 | 0.1885 | 0.1280 |
| 9600 | 0.1850 | 0.1280 |
| 9900 | 0.1760 | 0.1280 |
| 10200 | 0.1850 | 0.1280 |
| 10500 | 0.1938 | 0.1280 |
| 10800 | 0.1944 | 0.1280 |
| 11100 | 0.1853 | 0.1280 |
| 11400 | 0.1894 | 0.1280 |
| 11700 | 0.1935 | 0.1280 |
| 12000 | 0.1932 | 0.1280 |
| 12300 | 0.1988 | 0.1280 |
| 12600 | 0.1915 | 0.1280 |
| 12900 | 0.1847 | 0.1280 |
| 13200 | 0.1903 | 0.1280 |
| 13500 | 0.1880 | 0.1280 |
| 13800 | 0.1970 | 0.1280 |
| 14100 | 0.1994 | 0.1280 |
| 14400 | 0.1926 | 0.1280 |
| 14700 | 0.1967 | 0.1280 |
| 15000 | 0.1999 | 0.1280 |
| 15300 | 0.1988 | 0.1280 |
| 15600 | 0.2011 | 0.1280 |
| 15900 | 0.2026 | 0.1280 |
| 16200 | 0.2023 | 0.1280 |
| 16500 | 0.2005 | 0.1280 |
| 16800 | 0.2032 | 0.1280 |
| 17100 | 0.1964 | 0.1280 |
| 17400 | 0.2011 | 0.1280 |
| 17700 | 0.2008 | 0.1280 |
| 18000 | 0.1994 | 0.1280 |
Measured result
Verbatim from the run's summary.json.
checkpoint_round_trip
true
checkpoint_sha256
6a847ce6179567b0…checkpoint_source
None
checkpoint_source_sha256
None
checks
all_valid
true
beats_drop_baseline
false
beats_marginal_baseline
true
median_recovery
false
config_sha256
792fcbd457542727…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
17.5811209439528
median_bins_off
11.0
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
6.352187633514404
gpu
NVIDIA GeForce RTX 4080
held_out_nll_per_masked_token
3.593099948154908
icon_presentations
288000
initial_held_out_nll
6.218960549879478
inpainting
all_valid
true
attempted
56
close_reproduction_rate
0.0
decoding
chain_order
true
greedy_for_criteria
true
iterations
8
samples
1
drop
median_rgba_mae
0.005432628298232873
exact_reproduction_rate
0.0
icons
56
identity_policy
join_span
inpaint_seed
3501
marginal
helped
4
mean_recovery
-9243201.564606853
median_recovery
-1.5514862525194055
median_rgba_mae
0.02051230407286371
model
helped
22
mean_recovery
-2958359.976244374
median_recovery
-0.044644347957101746
median_rgba_mae
0.004989428255870249
paired
drop_minus_model
interval
-0.0022500233027615414, 0.0018352586314609215
interval_excludes_zero_above
false
mean
-0.0002073823356503099
n
56
positive
22
marginal_minus_model
interval
0.004811872256668611, 0.015588885461542124
interval_excludes_zero_above
true
mean
0.010200378859105368
n
56
positive
49
rows_sha256
6c4fed0fbf6ea80f…sheet_sha256
a3f65aeadbaf6096…span_length
4
task
span
last_train_loss
3.5865390300750732
loss_reduction_factor
1.7308064455799441
marginal_masked_token_accuracy
0.12803273896521486
marginal_nll_per_masked_token
4.7650980387679045
mask_mixture
drawn
span
288000
max_paths
2
random_rate
0.05, 0.95
span_max
4
weights
span
1.0
masked_token_accuracy
0.20111078631978954
metric_head
true
metrics_sha256
eeca047ef8232e16…model_parameters
531892
nll_ratio_to_marginal
0.7540453352527378
path_binding
false
predeclared_outcome
falsified
schema_version
1
selected_step
15600
sequence_length
1376
start_features
true
steps_run
18000
study_version
masked-span-l21-specialist-4-long
timing
peak_vram_gib
1.9460439682006836
prepare_seconds
102.04778214800172
train_seconds
1199.1558452319878
torch_version
2.14.0a0+4fdf77b940.nv26.08
train_icons
2681
trained_here
true
vocabulary
418
Visual output
State transitions
- planned2026-09-21T10:34:28Z
- completed2026-09-21T11:07:07Ztrained three times longer, the specialist's held-out likelihood keeps improving - 3.593, 0.754 of the floor, selected at step 15,600 - and its continuity reaches 17.6 bins against the copy policy's 25.6, and the edit still ties the join: mean -0.0002, interval [-0.0023, +0.0018] spanning zero, 22 of 56 helped, indistinguishable span for span from arm 20 (24 better, 32 worse); the marginal policy beaten on 49 of 56; 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-span-l21-specialist-4-long-5a50b51-2681icons-9b9b1699
config
configs/learning/masked-span-l21-specialist-4-long.yaml
outputs
checkpoint_root
data/processed/masked-span-l21-specialist-4-long
report_root
reports/learning/masked-span-l21-specialist-4-long