masked-overfit-l3-binding-438a98f-4icons-9b9b1699
completed —
Measured behaviour
modelposition-marginal floor
Table view
| optimizer step | model | position-marginal floor |
|---|---|---|
| 0 | 5.8277 | 4.9159 |
| 100 | 4.8832 | 4.9159 |
| 200 | 4.1721 | 4.9159 |
| 300 | 3.1250 | 4.9159 |
| 400 | 2.4681 | 4.9159 |
| 500 | 2.2187 | 4.9159 |
| 600 | 2.1464 | 4.9159 |
| 700 | 2.1197 | 4.9159 |
| 800 | 2.0984 | 4.9159 |
| 900 | 2.1029 | 4.9159 |
| 1000 | 2.0678 | 4.9159 |
| 1100 | 2.0555 | 4.9159 |
| 1200 | 2.0481 | 4.9159 |
| 1300 | 2.0384 | 4.9159 |
| 1400 | 2.0353 | 4.9159 |
| 1500 | 2.0288 | 4.9159 |
| 1600 | 2.0125 | 4.9159 |
| 1700 | 2.0072 | 4.9159 |
| 1800 | 2.0044 | 4.9159 |
| 1900 | 2.0093 | 4.9159 |
| 2000 | 1.9964 | 4.9159 |
| 2100 | 1.9965 | 4.9159 |
| 2200 | 1.9938 | 4.9159 |
| 2300 | 1.9929 | 4.9159 |
| 2400 | 1.9932 | 4.9159 |
| 2500 | 1.9902 | 4.9159 |
| 2600 | 1.9909 | 4.9159 |
| 2700 | 1.9813 | 4.9159 |
| 2800 | 1.9807 | 4.9159 |
| 2900 | 1.9897 | 4.9159 |
| 3000 | 1.9806 | 4.9159 |
modelposition-marginal floor
Table view
| optimizer step | model | position-marginal floor |
|---|---|---|
| 0 | 0.0140 | 0.5470 |
| 100 | 0.0701 | 0.5470 |
| 200 | 0.1473 | 0.5470 |
| 300 | 0.3100 | 0.5470 |
| 400 | 0.4432 | 0.5470 |
| 500 | 0.6255 | 0.5470 |
| 600 | 0.7588 | 0.5470 |
| 700 | 0.8191 | 0.5470 |
| 800 | 0.8527 | 0.5470 |
| 900 | 0.8640 | 0.5470 |
| 1000 | 0.9032 | 0.5470 |
| 1100 | 0.9453 | 0.5470 |
| 1200 | 0.9481 | 0.5470 |
| 1300 | 0.9663 | 0.5470 |
| 1400 | 0.9790 | 0.5470 |
| 1500 | 0.9719 | 0.5470 |
| 1600 | 0.9916 | 0.5470 |
| 1700 | 0.9972 | 0.5470 |
| 1800 | 0.9916 | 0.5470 |
| 1900 | 0.9818 | 0.5470 |
| 2000 | 0.9958 | 0.5470 |
| 2100 | 0.9972 | 0.5470 |
| 2200 | 0.9986 | 0.5470 |
| 2300 | 0.9958 | 0.5470 |
| 2400 | 0.9972 | 0.5470 |
| 2500 | 0.9986 | 0.5470 |
| 2600 | 1.0000 | 0.5470 |
| 2700 | 1.0000 | 0.5470 |
| 2800 | 1.0000 | 0.5470 |
| 2900 | 0.9972 | 0.5470 |
| 3000 | 1.0000 | 0.5470 |
Measured result
Verbatim from the run's summary.json.
checkpoint_round_trip
true
checkpoint_sha256
b5e756d8d74e2ca1…checks
all_valid
true
exact_path_reproduction
true
masked_token_accuracy
true
config_sha256
d9090bc6d3be8548…coordinate_tau
1.0
criteria
min_exact_path_reproduction_rate
1.0
min_masked_token_accuracy
0.99
require_all_valid
true
deterministic_algorithms
true
device
cuda
evaluation_icons
4
evaluation_split
train
first_train_loss
4.911922931671143
gpu
NVIDIA GeForce RTX 4080
held_out_nll_per_masked_token
1.98063345075824
icon_presentations
12000
initial_held_out_nll
5.827729170319074
inpainting
all_valid
true
attempted
4
decoding
greedy_for_criteria
true
iterations
8
samples
1
drop
median_rgba_mae
0.0034825291999515855
exact_reproduction_rate
1.0
icons
4
marginal
helped
0
mean_recovery
-1.8921178315003433
median_recovery
-2.0224896560213343
median_rgba_mae
0.010612782921810698
model
helped
4
mean_recovery
1.0
median_recovery
1.0
median_rgba_mae
0.0
paired
drop_minus_model
interval
-0.01649456796035647, 0.03891612145829885
interval_excludes_zero_above
false
mean
0.011210776748971191
n
4
positive
4
marginal_minus_model
interval
-0.0119139906416794, 0.04538188204206671
interval_excludes_zero_above
false
mean
0.016733945700193657
n
4
positive
4
rows_sha256
71d41d0a95011162…sheet_sha256
7f1f2ba48d25ddd0…last_train_loss
1.940286636352539
loss_reduction_factor
2.942356228553073
marginal_masked_token_accuracy
0.5469845722300141
marginal_nll_per_masked_token
4.9158920138280155
mask_mixture
drawn
geometry
1203
path
3603
random
3605
span
2422
style
1167
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
1.0
metrics_sha256
f552b07359c21059…model_parameters
542050
nll_ratio_to_marginal
0.4029041820257391
path_binding
true
predeclared_outcome
passed
schema_version
1
selected_step
3000
sequence_length
1376
steps_run
3000
study_version
masked-overfit-l3-binding
timing
peak_vram_gib
0.49772024154663086
prepare_seconds
0.2604321110120509
train_seconds
112.19813616099418
torch_version
2.14.0a0+4fdf77b940.nv26.08
train_icons
4
vocabulary
418
Visual output
State transitions
- planned2026-09-21T08:32:47Z
- completed2026-09-21T08:35:58Zthe bound model memorises four icons exactly: masked-token accuracy 1.000, 4 of 4 masked whole paths reproduced token for token with pixel-identical renders, every completion valid; memorisation reaches 0.992 by step 1,800 against l1's 0.999, so binding costs nothing here
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-overfit-l3-binding-438a98f-4icons-9b9b1699
config
configs/learning/masked-overfit-l3-binding.yaml
outputs
checkpoint_root
data/processed/masked-overfit-l3-binding
report_root
reports/learning/masked-overfit-l3-binding