CITADEL
StandbyNothing is live right now.
Eval accuracylast measured Tue 10 Nov, 08:59 UTC
%
0.561573 of 130 eval, 57 wrong
▲ +5.38 pts across 2 evals▲ +5.38 pts on the eval before
0%25%50%75%100%0%25%50%75%100%0%25%50%75%100%

Claviger: OneJev 0.8B (backfilled)

finishedRadeon 8060S

A small model that reads one Discord message and decides whether an agent should reply, and which one. 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, 08:40 UTC
ran for
17:53
config
learning rate 0.00005, batch 8, seed 0

run4/6step150/150epoch5/5config4/4seed0lr0.00005batch8modelOneJev1.6-0.8B 0.8Brevc89933citems479+130

Loss train 1.0680 ▼ −0.2701 eval 1.0154 lowest 0.7107 at step 17 run before, run 3 ▲ +1.0669 step 150
Timebase
Range
Smoothing
Reference
150 readings, 5 min/div5 min/div2 min/div
0123400:0005:0010:0015:00phase starting, 0.0 s to 7.2 sphase training, 7.2 s to 17:53epoch 1, 7.2 sepoch 3, 07:16epoch 4, 10:51epoch 5, 14:26trainingE3E4E51.3381Lowest last train loss across this project's 6 runs, run 5lowest last 0.0000, run 51.06800123400:0005:0010:0015:00phase starting, 0.0 s to 7.2 sphase training, 7.2 s to 17:53epoch 1, 7.2 sepoch 2, 03:41epoch 3, 07:16epoch 4, 10:51epoch 5, 14:26trainingepoch 2epoch 3epoch 4epoch 51.3381Lowest last train loss across this project's 6 runs, run 5lowest last 0.0000, run 5eval 1.01541.0680▼ −0.27010123400:0002:0004:0006:0008:0010:0012:0014:0016:00phase starting, 0.0 s to 7.2 sphase training, 7.2 s to 17:53epoch 1, 7.2 sepoch 2, 03:41epoch 3, 07:16epoch 4, 10:51epoch 5, 14:26trainingepoch 2epoch 3epoch 4epoch 51.3381Lowest last train loss across this project's 6 runs, run 5lowest last 0.0000, run 5eval 1.01541.0680▼ −0.2701

Eval measures

Eval losslower is better
1.0154▼ −0.0472 since first
02 →
ECEcalibration error, lower is better
0.2027▼ −0.0009 since first
01

Elapsed

214.7 s

epoch interval

150 of 150 steps, 5 of 5 epochs, 100%

7.16 s per step, 8.4 steps a minute

Eval accuracy by eval

+0.0538

▲ on the eval before

eval 1 0.5077, eval 2 0.5615

Against the other runs

6 of 6

by last eval accuracy

best so far 0.8615, run 6

▼ −0.2154 against run 3

Train minus eval

+0.0526

step 150

train 1.0680, eval 1.0154

Set-up

479 train, 130 eval

synthetic, learning rate 0.00005, batch size 8, grad accum 2, epochs 5, optimizer adamw, seed 0

Learning rate

0.00e+0

▼ −1.25e-5 since first

On one axis

phase starting, 0.0 s to 7.2 sphase training, 7.2 s to 17:53E2E3E4E5traininglowest last 0.0000Train loss1.0680▼ −0.27014.000.001.33811.0680last ▼ −0.2701best 0.8615, run 6Last-value range, 6 runs, 0.5615 to 0.8615last-value range, n 6Eval accuracy0.5615▲ +0.05381.000.400.50770.5615last ▲ +0.0538Learning rate0.00e+0▼ −1.25e-56.0e-50.0e+01.25e-50.00e+0last ▼ −1.25e-500:0002:0004:0006:0008:0010:0012:0014:0016:00

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

4 of 6 finished, 0.5615, config 4 of 4 (learning rate 0.00005, batch 8), seed 0, Tue 10 Nov, 08:40 UTC

Runs as a table
StartedStatusEval accuracyStepsConfig
Tue 10 Nov, 09:19 UTC| finished0.8615150 of 150learning rate 0.000005, batch 8, seed 2 of 3
Tue 10 Nov, 08:59 UTC| finished0.8154150 of 150learning rate 0.000005, batch 8, seed 1 of 3
Tue 10 Nov, 08:40 UTC| finished0.5615150 of 150learning rate 0.00005, batch 8, seed 0
Tue 10 Nov, 08:21 UTC| finished0.7769150 of 150learning rate 0.00002, batch 8, seed 0
Tue 10 Nov, 08:01 UTC| finished0.7923150 of 150learning rate 0.00001, batch 8, seed 0
Tue 10 Nov, 07:42 UTC| finished0.8000150 of 150learning rate 0.000005, batch 8, seed 0 of 3

Plate

Model
OmniJev/OneJev1.6-0.8B, 0.8B parameters revision c89933c
Data
Synthetic Discord triage set, mixed rosters, synthetic479 train130 eval
Config
learning rate 0.00005, batch size 8, grad accum 2, epochs 5, optimizer adamw, seed 0
Hardware
Radeon 8060S, ROCm

Log

Since startMark
0.0 sphase starting
0.0 seval 1
7.2 sphase training
7.2 sepoch 1
221.9 sepoch 2
436.6 sepoch 3
651.4 sepoch 4
866.1 sepoch 5
1073.7 seval 2
196 h 11 min laterwent public

Readings

151 rows as a table

rows 1 to 100 of 151

StepSecondsTrain lossEval lossEval accuracyECELearning rate
00.0-1.06260.50770.2036-
17.21.3381---1.25e-5
214.30.7801---2.50e-5
321.51.9680---3.75e-5
428.60.8810---5.00e-5
535.81.1864---5.00e-5
642.91.2915---5.00e-5
750.11.1936---4.99e-5
857.31.2785---4.99e-5
964.41.1431---4.99e-5
1071.61.1456---4.98e-5
1178.71.0765---4.97e-5
1285.90.9600---4.96e-5
1393.10.9566---4.95e-5
14100.21.0249---4.94e-5
15107.40.9193---4.93e-5
16114.50.7298---4.92e-5
17121.70.7107---4.90e-5
18128.80.9571---4.89e-5
19136.01.8223---4.87e-5
20143.21.0655---4.85e-5
21150.31.1273---4.83e-5
22157.51.4168---4.81e-5
23164.61.2954---4.79e-5
24171.81.0946---4.77e-5
25178.91.2146---4.75e-5
26186.11.3026---4.73e-5
27193.31.0554---4.70e-5
28200.41.4847---4.67e-5
29207.61.4038---4.65e-5
30214.71.0023---4.62e-5
31221.91.1285---4.59e-5
32229.11.2284---4.56e-5
33236.21.1261---4.53e-5
34243.41.0221---4.50e-5
35250.51.0590---4.46e-5
36257.71.0715---4.43e-5
37264.81.0522---4.40e-5
38272.01.0614---4.36e-5
39279.21.0404---4.32e-5
40286.31.1076---4.29e-5
41293.51.0251---4.25e-5
42300.61.0761---4.21e-5
43307.81.3600---4.17e-5
44315.01.1117---4.13e-5
45322.11.3098---4.09e-5
46329.31.1550---4.05e-5
47336.41.0991---4.00e-5
48343.61.0326---3.96e-5
49350.71.1040---3.92e-5
50357.91.0517---3.87e-5
51365.11.1977---3.83e-5
52372.21.4462---3.78e-5
53379.41.2123---3.73e-5
54386.51.1943---3.69e-5
55393.71.1488---3.64e-5
56400.81.0566---3.59e-5
57408.01.0474---3.54e-5
58415.21.1010---3.49e-5
59422.31.1371---3.44e-5
60429.51.1744---3.39e-5
61436.61.1062---3.34e-5
62443.81.0627---3.29e-5
63451.01.0816---3.24e-5
64458.11.0559---3.19e-5
65465.31.0254---3.14e-5
66472.41.1452---3.09e-5
67479.61.0871---3.03e-5
68486.70.9441---2.98e-5
69493.91.0893---2.93e-5
70501.11.1063---2.88e-5
71508.21.1082---2.82e-5
72515.41.0825---2.77e-5
73522.51.0808---2.71e-5
74529.71.0500---2.66e-5
75536.91.0134---2.61e-5
76544.01.0705---2.55e-5
77551.21.1897---2.50e-5
78558.31.0066---2.45e-5
79565.51.1035---2.39e-5
80572.60.9848---2.34e-5
81579.81.0956---2.29e-5
82587.01.1189---2.23e-5
83594.11.0741---2.18e-5
84601.31.1619---2.12e-5
85608.41.0705---2.07e-5
86615.61.0920---2.02e-5
87622.71.1724---1.97e-5
88629.91.0156---1.91e-5
89637.13.5717---1.86e-5
90644.21.2913---1.81e-5
91651.41.0849---1.76e-5
92658.51.0150---1.71e-5
93665.70.9716---1.66e-5
94672.91.1821---1.61e-5
95680.01.0782---1.56e-5
96687.21.1550---1.51e-5
97694.31.1388---1.46e-5
98701.50.9739---1.41e-5
99708.61.2067---1.36e-5

Worked out

ReadoutValueWorked out from
Run complete100%150 steps over 150
Step time7.16 s1066.5 s over 149 steps, between readings. Startup and intervals that hold an eval are left out
Steps per minute8.460 s over the step time of 7.16 s
Epoch interval214.7 sMean of 4 gaps between epoch marks
Train loss since first reading−0.27011.0680 at step 150 minus 1.3381 at step 1
Lowest train loss0.7107Lowest of 150 readings, at step 17
Eval accuracy, as shown56.2%0.5615 times 100, to one decimal. Changes on that plate are percentage points
Eval accuracy since first eval+0.05380.5615 minus 0.5077
Eval loss since first eval−0.04721.0154 minus 1.0626
ECE since first eval−0.00090.2027 minus 0.2036
Eval change+0.0538Eval 2 at 0.5615 minus eval 1 at 0.5077
Train minus eval loss+0.05261.0680 minus 1.0154, both at step 150
Wrong answers57130 eval items minus 73 correct. 73 over 130 is the only count that gives 0.5615
Rank6 of 6By last eval accuracy, across this project's 6 runs with an eval
Best so far0.8615Highest last eval accuracy, run 6
Against the run before−0.21540.5615 minus 0.7769 on run 3
Lowest last train loss0.0000Lowest last train loss across this project's 6 runs, run 5
Items per second1.11 step over 7.16 s, times batch size 8
Train loss against run 3+1.06691.0680 minus 0.0010 on run 3, both at step 150
Learning rate since first reading−1.25e-50.00e+0 minus 1.25e-5