ar-capacity-i5-91328fa-3sizes-9b9b1699
completed —
Measured behaviour
modelposition-marginal floor
Table view
| optimizer step | model | position-marginal floor |
|---|---|---|
| 300 | 3.6612 | 3.9290 |
| 600 | 3.6480 | 3.9290 |
| 900 | 3.8384 | 3.9290 |
| 1200 | 3.9849 | 3.9290 |
| 1500 | 4.2191 | 3.9290 |
| 1800 | 4.6009 | 3.9290 |
| 2100 | 4.7426 | 3.9290 |
| 2400 | 5.0071 | 3.9290 |
| 2700 | 5.2550 | 3.9290 |
| 3000 | 5.4942 | 3.9290 |
| 300 | 3.6565 | 3.9290 |
| 600 | 3.9004 | 3.9290 |
| 900 | 4.5608 | 3.9290 |
| 1200 | 5.0955 | 3.9290 |
| 1500 | 5.5763 | 3.9290 |
| 1800 | 5.9573 | 3.9290 |
| 2100 | 6.1508 | 3.9290 |
| 2400 | 6.3368 | 3.9290 |
| 2700 | 6.5507 | 3.9290 |
| 300 | 3.6899 | 3.9290 |
| 600 | 4.4713 | 3.9290 |
| 900 | 5.0386 | 3.9290 |
| 1200 | 5.2441 | 3.9290 |
| 1500 | 5.4002 | 3.9290 |
| 1800 | 5.7515 | 3.9290 |
| 2100 | 5.8451 | 3.9290 |
| 2400 | 6.0244 | 3.9290 |
| 2700 | 6.2210 | 3.9290 |
Measured result
Verbatim from the run's summary.json.
arms
epochs_to_best
3.5807534502051475
fraction
1.0
held_out_nll
3.64796488139269
icons
2681
label
base-525k
marginal_nll
3.9289882210903344
parameters
524674
ratio_to_own_floor
0.9284743746012917
selected_step
600
steps_run
3000
trace
fraction
1.0
held_out_nll
3.6611640545050923
label
base-525k
marginal_nll
3.9289882210903344
step
300
train_loss
3.2960057258605957
fraction
1.0
held_out_nll
3.64796488139269
label
base-525k
marginal_nll
3.9289882210903344
step
600
train_loss
3.3200597763061523
fraction
1.0
held_out_nll
3.8384189372997466
label
base-525k
marginal_nll
3.9289882210903344
step
900
train_loss
2.903935194015503
fraction
1.0
held_out_nll
3.984894900990683
label
base-525k
marginal_nll
3.9289882210903344
step
1200
train_loss
1.603338360786438
fraction
1.0
held_out_nll
4.219101442814329
label
base-525k
marginal_nll
3.9289882210903344
step
1500
train_loss
2.4615190029144287
fraction
1.0
held_out_nll
4.600927306373807
label
base-525k
marginal_nll
3.9289882210903344
step
1800
train_loss
1.678338646888733
… and 4 more
train_seconds
305.51750963799714
epochs_to_best
1.7903767251025737
fraction
1.0
held_out_nll
3.6564721382704044
icons
2681
label
mid-1.6m
marginal_nll
3.9289882210903344
parameters
1614978
ratio_to_own_floor
0.9306396284526646
selected_step
300
steps_run
2700
trace
fraction
1.0
held_out_nll
3.6564721382704044
label
mid-1.6m
marginal_nll
3.9289882210903344
step
300
train_loss
2.8141608238220215
fraction
1.0
held_out_nll
3.900353470094477
label
mid-1.6m
marginal_nll
3.9289882210903344
step
600
train_loss
2.686300277709961
fraction
1.0
held_out_nll
4.560797911530256
label
mid-1.6m
marginal_nll
3.9289882210903344
step
900
train_loss
2.0599172115325928
fraction
1.0
held_out_nll
5.095544142930674
label
mid-1.6m
marginal_nll
3.9289882210903344
step
1200
train_loss
0.7050114274024963
fraction
1.0
held_out_nll
5.5762640811500255
label
mid-1.6m
marginal_nll
3.9289882210903344
step
1500
train_loss
1.7525978088378906
fraction
1.0
held_out_nll
5.957315375573159
label
mid-1.6m
marginal_nll
3.9289882210903344
step
1800
train_loss
0.793700098991394
… and 3 more
train_seconds
423.2090208479931
epochs_to_best
1.7903767251025737
fraction
1.0
held_out_nll
3.6899182573042997
icons
2681
label
large-4.8m
marginal_nll
3.9289882210903344
parameters
4817570
ratio_to_own_floor
0.9391522828949357
selected_step
300
steps_run
2700
trace
fraction
1.0
held_out_nll
3.6899182573042997
label
large-4.8m
marginal_nll
3.9289882210903344
step
300
train_loss
2.3863632678985596
fraction
1.0
held_out_nll
4.471346667475041
label
large-4.8m
marginal_nll
3.9289882210903344
step
600
train_loss
1.5283855199813843
fraction
1.0
held_out_nll
5.038643451437789
label
large-4.8m
marginal_nll
3.9289882210903344
step
900
train_loss
1.4836143255233765
fraction
1.0
held_out_nll
5.24414944676191
label
large-4.8m
marginal_nll
3.9289882210903344
step
1200
train_loss
0.5497051477432251
fraction
1.0
held_out_nll
5.400222925269124
label
large-4.8m
marginal_nll
3.9289882210903344
step
1500
train_loss
1.4778703451156616
fraction
1.0
held_out_nll
5.7514857143849705
label
large-4.8m
marginal_nll
3.9289882210903344
step
1800
train_loss
0.6317853927612305
… and 3 more
train_seconds
599.000456601003
axis
capacity
best_ratio_to_own_floor
0.9284743746012917
checks
best_arm_clears_magnitude_bar
false
monotone_in_capacity
false
second_step_still_buys
false
config_sha256
88823e0693e79f6e…criteria
max_best_ratio
0.85
min_second_doubling_share
0.25
deterministic_algorithms
true
device
cuda
gpu
NVIDIA GeForce RTX 4080
improvement_first_doubling
-0.008507256877714475
improvement_second_doubling
-0.03344611903389527
predeclared_outcome
not_limited_by_capacity
prepare_seconds
102.16331128599995
schema_version
1
second_doubling_share
0.0
study_version
ar-capacity-i5
torch_version
2.14.0a0+4fdf77b940.nv26.08
validation_icons
339
State transitions
- planned2026-09-21T12:55:00Z
- completed2026-09-21T13:40:00Zcapacity is not a lever and is mildly harmful; 9x the parameters is monotonically worse and every arm peaks within two epochs
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
ar-capacity-i5-91328fa-3sizes-9b9b1699
config
configs/learning/ar-capacity-i5.yaml