CITADEL
StandbyNothing is live right now.
Eval accuracylast measured Tue 10 Nov, 16:24 UTC
%
0.0847
one eval
0%25%50%75%100%0%25%50%75%100%0%25%50%75%100%

Claviger router: OneJev 0.8B (backfilled)

finishedRadeon 8060S

A model that routes a request to the one destination, of many, that should take it. Backfilled: recorded after the fact from the run's own logs and records. Times inside a run are estimated, and there are no GPU sensor readings.

started
Tue 10 Nov, 15:27 UTC
ran for
46:18
config
learning rate 0.00005, batch 1, seed 0

run3/5step42/42epoch1/1config3/3seed0lr0.00005batch1modelOneJev1.6-0.8B 0.8Brevc89933c

Loss train 3.3927 ▲ +3.3052 eval 3.5070 lowest 0.0181 at step 2 run before, run 2 ▲ +2.7765 step 42
Timebase
Range
Smoothing
Reference
42 readings, 30 min/div10 min/div5 min/div
024600:0030:00phase starting, 0.0 s to 66.2 sphase training, 66.2 s to 46:18epoch 1, 66.2 straining+3.3052since first0.0876Lowest last train loss across this project's 5 runs, run 4lowest last 0.0014, run 43.3927024600:0010:0020:0030:0040:00phase starting, 0.0 s to 66.2 sphase training, 66.2 s to 46:18epoch 1, 66.2 straining0.0876Lowest last train loss across this project's 5 runs, run 4lowest last 0.0014, run 4eval 3.50703.3927▲ +3.3052024600:0005:0010:0015:0020:0025:0030:0035:0040:0045:00phase starting, 0.0 s to 66.2 sphase training, 66.2 s to 46:18epoch 1, 66.2 straining0.0876Lowest last train loss across this project's 5 runs, run 4lowest last 0.0014, run 4eval 3.50703.3927▲ +3.3052

Eval measures

Eval losslower is better
3.5070
04 →
ECEcalibration error, lower is better
0.0146
01

Elapsed

66.2 s to epoch 1, 2712.4 s to eval 1

42 of 42 steps, 1 of 1 epochs, 100%

66.2 s per step, 0.9 steps a minute

Against the other runs

3 of 3

by last eval accuracy

best so far 0.9233, run 1

▼ −0.8280 against run 2

Train minus eval

−0.1143

step 42

train 3.3927, eval 3.5070

Set-up

666 train

synthetic, learning rate 0.00005, batch size 1, grad accum 16, epochs 1, optimizer adamw, seed 0

Learning rate

0.00e+0

▼ −5.00e-5 since first

Input

  • Runs128 runs
  • Agentsnone showing
  • Treesnone showing
  • Projectsnone showing

Check

Every entry is checked on its way out; some are held for review.

Not shown on any run: data samples, predictions, raw logs, file paths, machine names

Delay

2 min, set

Watching the agents work, about 2 min behind.

One way

Nothing here reaches the agents.

This page

This page updates by itself.

visits counted with Google Analytics · choose

Runsheight is eval accuracyall recordings ▸ | finished× failed■ stopped○ lost● live

3 of 5 finished, 0.0847, config 3 of 3 (learning rate 0.00005, batch 1), seed 0, Tue 10 Nov, 15:27 UTC

Runs as a table
StartedStatusEval accuracyStepsConfig
Tue 10 Nov, 17:10 UTC| finished-42 of 42learning rate 0.00001, batch 1, seed 2 of 3
Tue 10 Nov, 16:24 UTC| finished-42 of 42learning rate 0.00001, batch 1, seed 1 of 3
Tue 10 Nov, 15:27 UTC| finished0.084742 of 42learning rate 0.00005, batch 1, seed 0
Tue 10 Nov, 14:30 UTC| finished0.912742 of 42learning rate 0.00002, batch 1, seed 0
Tue 10 Nov, 13:33 UTC| finished0.923342 of 42learning rate 0.00001, batch 1, seed 0 of 3

Plate

Model
OmniJev/OneJev1.6-0.8B, 0.8B parameters revision c89933c
Data
Synthetic routing worlds, synthetic
Config
learning rate 0.00005, batch size 1, grad accum 16, epochs 1, optimizer adamw, seed 0
Hardware
Radeon 8060S, ROCm

Log

Since startMark
0.0 sphase starting
66.2 sphase training
66.2 sepoch 1
2778.6 seval 1
189 h 25 min laterwent public

Readings

42 rows as a table
StepSecondsTrain lossEval lossEval accuracyECELearning rate
166.20.0876---5.00e-5
2132.30.0181---4.99e-5
3198.50.0327---4.97e-5
4264.64.4199---4.93e-5
5330.84.3565---4.88e-5
6396.93.4358---4.82e-5
7463.14.1156---4.74e-5
8529.33.6576---4.65e-5
9595.44.4331---4.54e-5
10661.63.7818---4.43e-5
11727.72.9675---4.30e-5
12793.93.9769---4.16e-5
13860.03.8595---4.02e-5
14926.23.4561---3.86e-5
15992.43.5178---3.69e-5
161058.53.7054---3.52e-5
171124.73.3877---3.35e-5
181190.83.3363---3.16e-5
191257.03.8628---2.98e-5
201323.13.3338---2.79e-5
211389.33.5027---2.60e-5
221455.53.6891---2.40e-5
231521.64.2289---2.21e-5
241587.83.4918---2.02e-5
251653.93.4030---1.84e-5
261720.13.2864---1.65e-5
271786.23.5850---1.48e-5
281852.43.5730---1.31e-5
291918.63.7645---1.14e-5
301984.73.0648---9.84e-6
312050.93.4624---8.37e-6
322117.03.0825---6.99e-6
332183.23.8877---5.71e-6
342249.33.1491---4.55e-6
352315.53.5161---3.51e-6
362381.73.6975---2.60e-6
372447.83.9955---1.81e-6
382514.03.5585---1.17e-6
392580.12.9712---6.58e-7
402646.33.9065---2.93e-7
412712.42.8897---7.34e-8
422778.63.39273.50700.08470.01460.00e+0

Worked out

ReadoutValueWorked out from
Run complete100%42 steps over 42
Step time66.2 s2646.2 s over 40 steps, between readings. Startup and intervals that hold an eval are left out
Steps per minute0.960 s over the step time of 66.16 s
Train loss since first reading+3.30523.3927 at step 42 minus 0.0876 at step 1
Lowest train loss0.0181Lowest of 42 readings, at step 2
Eval accuracy, as shown8.5%0.0847 times 100, to one decimal. Changes on that plate are percentage points
Train minus eval loss−0.11433.3927 minus 3.5070, both at step 42
Rank3 of 3By last eval accuracy, across this project's 3 runs with an eval
Best so far0.9233Highest last eval accuracy, run 1
Against the run before−0.82800.0847 minus 0.9127 on run 2
Lowest last train loss0.0014Lowest last train loss across this project's 5 runs, run 4
Items per second0.01 step over 66.16 s, times batch size 1
Train loss against run 2+2.77653.3927 minus 0.6162 on run 2, both at step 42
Learning rate since first reading−5.00e-50.00e+0 minus 5.00e-5