MojiDiff

← experiments

ar-corpus-i3-metric-29f9715-2681icons-9b9b1699

failed —

Measured behaviour

Held-out negative log likelihood per free token
modelposition-marginal floor
nats per tokenoptimizer step02465001000150020002500selected
Table view
optimizer stepmodelposition-marginal floor
3003.69463.9290
6003.69583.9290
9003.85773.9290
12004.05643.9290
15004.36363.9290
18004.50623.9290
21004.69863.9290
24005.19943.9290
27005.36493.9290

Measured result

Verbatim from the run's summary.json.

bits_per_free_token
5.330138705054989
checkpoint_round_trip
true
checkpoint_sha256
f89357e009c04835…
checks
all_valid
true
distinct_sample_rate
true
nll_ratio_to_marginal
false
config_sha256
4439ade27fbda1ab…
criteria
max_nll_ratio_to_marginal
0.5
min_distinct_sample_rate
0.9
require_all_valid
true
deterministic_algorithms
true
device
cuda
gpu
NVIDIA GeForce RTX 4080
held_out_nll_per_free_token
3.694570615402304
icon_presentations
100800
marginal_nll_per_free_token
3.9289882210903344
metrics_sha256
6619f3a1af30225e…
model_parameters
526498
nll_ratio_to_marginal
0.9403363938762389
predeclared_outcome
falsified
samples
all_valid
true
corpus_ink_coverage_iqr
0.2606095679012346, 0.3869116512345679
count
32
digests
aaf9131a7bf79fedb0c43790b7a60c856b626889912af3bfe6b8de0e87daf3f9, 7d634deabf395bfd7cb85ad3584a56844b0453fdb6bb9ee3337a7327dd1c9b88, d2ff606707936362039b84ecb8389e6181709c006dd9e9f6ac4356842449731f, 58c06b2287f7219ec2ac621ad58509255ddcfafcb1414d8280017c105d9d650c, 657ed08e33c13225e47f2cad1463d1655b71b224589434809335c08f80f44a58, fd1fb3c26f38be63af7c68b12434065f1aa9c138777a91c71a6b9a12f5d69aac … and 26 more
distinct
32
distinct_rate
1.0
median_ink_coverage
0.4496527777777778
median_inside_corpus_iqr
false
reference_icons
128
schema_version
1
selected_step
300
sequence_length
1376
study_version
ar-corpus-i3-metric
timing
argmax_divergences
0
cache_speedup
1.134061340696624
cached_decode_seconds
1.6576374430005671
cached_forward_calls
1376
peak_vram_gib
3.6308536529541016
prepare_seconds
102.05215593600587
seconds_per_step
0.10220026211555469
smallest_divergence_margin
None
train_seconds
275.9407077119977
uncached_decode_seconds
1.8798625409981469
torch_version
2.14.0a0+4fdf77b940.nv26.08
train_icons
2681
validation_icons
339
vocabulary
418

State transitions

  1. planned2026-09-21T11:05:00Z
  2. failed2026-09-21T11:25:00Zthe uncached latency diagnostic never passed the segment kinds, so it raised under metric coordinates after training had completed; nothing was written

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-corpus-i3-metric-29f9715-2681icons-9b9b1699
config
configs/learning/ar-corpus-i3-metric.yaml