MojiDiff

← experiments

masked-overfit-l1-f51c119-4icons-9b9b1699

completed —

Measured behaviour

Held-out negative log likelihood per free token
modelposition-marginal floor
nats per tokenoptimizer step02460100020003000selected
Table view
optimizer stepmodelposition-marginal floor
05.78084.9159
1004.83174.9159
2003.67154.9159
3002.46384.9159
4002.21524.9159
5002.16954.9159
6002.12464.9159
7002.09974.9159
8002.08444.9159
9002.06584.9159
10002.05004.9159
11002.04014.9159
12002.03204.9159
13002.02444.9159
14002.02004.9159
15002.00764.9159
16002.00144.9159
17001.99654.9159
18001.99284.9159
19001.99694.9159
20001.98584.9159
21001.98824.9159
22001.97824.9159
23001.97984.9159
24001.97764.9159
25001.97884.9159
26001.97884.9159
27001.97034.9159
28001.97294.9159
29001.97484.9159
30001.96984.9159
Masked-token accuracy over unforced positions
modelposition-marginal floor
accuracyoptimizer step00.250.50.7510100020003000selected
Table view
optimizer stepmodelposition-marginal floor
00.02380.5470
1000.07150.5470
2000.20620.5470
3000.44180.5470
4000.62410.5470
5000.73630.5470
6000.82050.5470
7000.87940.5470
8000.90460.5470
9000.94670.5470
10000.96210.5470
11000.97760.5470
12000.98320.5470
13000.99160.5470
14000.99160.5470
15000.99720.5470
16000.99720.5470
17000.99720.5470
18000.99860.5470
19000.98880.5470
20000.99860.5470
21001.00000.5470
22001.00000.5470
23000.99720.5470
24001.00000.5470
25000.99860.5470
26001.00000.5470
27001.00000.5470
28001.00000.5470
29000.99860.5470
30001.00000.5470

Measured result

Verbatim from the run's summary.json.

checkpoint_round_trip
true
checkpoint_sha256
d67d6ae3d447d38e…
checks
all_valid
true
exact_path_reproduction
true
loss_reduction
false
masked_token_accuracy
true
config_sha256
64bf0b75e86afc24…
coordinate_tau
1.0
criteria
min_exact_path_reproduction_rate
1.0
min_loss_reduction_factor
100.0
min_masked_token_accuracy
0.99
require_all_valid
true
deterministic_algorithms
true
device
cuda
evaluation_icons
4
evaluation_split
train
first_train_loss
4.840734004974365
gpu
NVIDIA GeForce RTX 4080
held_out_nll_per_masked_token
1.9697834030943198
icon_presentations
12000
initial_held_out_nll
5.7808438979225105
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
b737fc98c371338e…
sheet_sha256
0f12c5ea2b77f8ee…
last_train_loss
1.938839077949524
loss_reduction_factor
2.9347611970135503
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
b91e20d2f8b79b6d…
model_parameters
526498
nll_ratio_to_marginal
0.4006970449215473
predeclared_outcome
falsified
schema_version
1
selected_step
3000
sequence_length
1376
steps_run
3000
study_version
masked-overfit-l1
timing
peak_vram_gib
0.4974055290222168
prepare_seconds
0.25998882099520415
train_seconds
106.27825578599004
torch_version
2.14.0a0+4fdf77b940.nv26.08
train_icons
4
vocabulary
418

Visual output

inpainting
inpainting.png

State transitions

  1. planned2026-09-21T08:12:54Z
  2. completed2026-09-21T08:16:16Zthe plumbing is sound: masked-token accuracy 1.000 over the free positions under the fixed masks, 4 of 4 masked whole paths reproduced token for token by greedy grammar-ordered completion with pixel-identical renders, every greedy, sampled and marginal completion valid; the loss-reduction check is falsified as written because under a soft coordinate target the exact-token likelihood cannot fall below the kernel's own entropy, about 2 nats per coordinate, so 2.9x is a ceiling set by the target rather than a fact about learning

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-l1-f51c119-4icons-9b9b1699
config
configs/learning/masked-overfit-l1.yaml
outputs
checkpoint_root
data/processed/masked-overfit-l1
report_root
reports/learning/masked-overfit-l1