CITADEL
StandbyNothing is live right now.
Eval accuracylast measured Sat 21 Nov, 18:26 UTC
%
0.8472377 of 445 eval, 68 wrong
▲ +2.92 pts across 3 evals▲ +0.45 pts on the eval before
0%25%50%75%100%0%25%50%75%100%0%25%50%75%100%

Familiar decider: Laya's encoder

finishedRadeon 8060S

One of the small models that decide whether Familiar, a desktop companion, reacts to something happening on the PC, and how. Every candidate trains on the same 6,000 decision points from synthetic timelines. The encoders also get one run on the full pool.

started
Sat 21 Nov, 17:47 UTC
ran for
25:53
config
learning rate 0.00005, batch 16, seed 0

run4/6step1032/1032epoch3/3config4/5seed0lr0.00005batch16modellaya 395Mrev7b928d8items5503+445

Loss train 0.4733 ▼ −1.3076 lowest 0.0441 at step 1030 run before, run 3 ▲ +0.2368 step 1032
Timebase
Range
Smoothing
Reference
104 readings, 10 min/div5 min/div5 min/div
012300:0010:0020:00phase starting, 0.0 s to 37.3 sphase training, 37.3 s to 25:53epoch 1, 09:02epoch 2, 17:28epoch 3, 25:37trainingepoch 1epoch 2−1.3076since first1.7809Lowest last train loss across this project's 6 runs, run 1lowest last 0.0241, run 10.4733012300:0005:0010:0015:0020:0025:00phase starting, 0.0 s to 37.3 sphase training, 37.3 s to 25:53epoch 1, 09:02epoch 2, 17:28epoch 3, 25:37trainingepoch 1epoch 21.7809Lowest last train loss across this project's 6 runs, run 1lowest last 0.0241, run 10.4733▼ −1.3076012300:0005:0010:0015:0020:0025:00phase starting, 0.0 s to 37.3 sphase training, 37.3 s to 25:53epoch 1, 09:02epoch 2, 17:28epoch 3, 25:37trainingepoch 1epoch 21.7809Lowest last train loss across this project's 6 runs, run 1lowest last 0.0241, run 10.4733▼ −1.3076

Eval measures

Elapsed

497.6 s

epoch interval

1032 of 1032 steps, 3 of 3 epochs, 100%

1.45 s per step, 41.3 steps a minute

Eval accuracy by eval

+0.0045

▲ on the eval before

eval 1 0.8180, eval 3 0.8472

Against the other runs

3 of 6

by last eval accuracy

best so far 0.9580, run 1

▼ −0.0427 against run 3

Set-up

5503 train, 445 eval

synthetic, learning rate 0.00005, batch size 16, epochs 3, max length 512, precision bf16-mixed, loss ce, optimizer adamw, scheduler linear, seed 0

Learning rate

0.00e+0

▼ −8.20e-6 since first

On one axis

phase starting, 0.0 s to 37.3 sphase training, 37.3 s to 25:53E1E2E3traininglowest last 0.0241Train loss0.4733▼ −1.30763.000.001.78090.4733last ▼ −1.3076best 0.9580, run 1Last-value range, 6 runs, 0.5618 to 0.9580Eval accuracy0.8472▲ +0.02921.000.400.81800.8472last ▲ +0.0292Learning rate0.00e+0▼ −8.20e-66.0e-50.0e+08.20e-6last 0.00e+0Power, whole chip, CPU and GPU118 W▲ +34 W140 W80 W84 Wlast 118 WHottest, one board zone93 °C▲ +1 °C98 °C88 °C92 °Clast 93 °C00:0005:0010:0015:0020:0025:00
GPUpower: whole chip, CPU and GPUtemperature: one board zone

GPU 0

Wwhole chip, CPU and GPU

118 W whole chip, CPU and GPU

use92%
memory, whole GPU2.1 GB

peak 23.7 GB, whole GPU

93 °C one board zone

history 0 to 124 W

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, 0.8472, config 4 of 5 (learning rate 0.00005, batch 16), seed 0, Sat 21 Nov, 17:47 UTC

Runs as a table
StartedStatusEval accuracyStepsConfig
Sat 21 Nov, 22:15 UTC| finished0.84721032 of 1032batch 16, seed 2 of 2
Sat 21 Nov, 21:36 UTC| finished0.64271032 of 1032batch 16, seed 1 of 2
Sat 21 Nov, 17:47 UTC| finished0.84721032 of 1032learning rate 0.00005, batch 16, seed 0
Sat 21 Nov, 17:08 UTC| finished0.88991032 of 1032learning rate 0.00002, batch 16, seed 0
Sat 21 Nov, 16:30 UTC| finished0.56181032 of 1032learning rate 0.0001, batch 16, seed 0
Sat 21 Nov, 08:25 UTC| finished0.95807274 of 7274batch 16, seed 0

Plate

Model
convaiinnovations/laya, 395M parameters revision 7b928d8
Data
Familiar decisions, shared draw, plain-case intents, synthetic5503 train445 eval
Config
learning rate 0.00005, batch size 16, epochs 3, max length 512, precision bf16-mixed, loss ce, optimizer adamw, scheduler linear, seed 0
Hardware
Radeon 8060S, ROCm

Log

Since startMark
0.0 sphase starting
7.0 swent public
37.3 sphase training
531.2 seval 1
542.6 sepoch 1
1043.2 seval 2
1048.3 sepoch 2
1537.9 sepoch 3
1553.3 seval 3

Readings

106 rows as a table

rows 1 to 100 of 106

StepSecondsTrain lossEval accuracyLearning rate
1037.31.7809-8.20e-6
2051.62.0503-1.64e-5
3065.82.3358-2.46e-5
4080.21.6692-3.28e-5
5094.51.0751-4.10e-5
60108.91.6479-4.92e-5
70123.31.3478-4.95e-5
80137.71.5180-4.90e-5
90151.91.7198-4.85e-5
100166.31.1063-4.80e-5
110180.51.3937-4.75e-5
120194.71.0262-4.70e-5
130209.01.0729-4.64e-5
140223.21.4564-4.59e-5
150237.51.1858-4.54e-5
160251.70.4917-4.49e-5
170266.11.0274-4.44e-5
180280.60.8411-4.39e-5
190295.01.0247-4.34e-5
200309.60.8056-4.28e-5
210323.90.9611-4.23e-5
220338.21.0833-4.18e-5
230352.60.9648-4.13e-5
240366.90.3406-4.08e-5
250381.20.6583-4.03e-5
260395.51.1163-3.98e-5
270409.80.6008-3.92e-5
280424.00.8083-3.87e-5
290438.20.9464-3.82e-5
300452.60.4327-3.77e-5
310467.01.1988-3.72e-5
320481.31.0851-3.67e-5
330495.80.5109-3.61e-5
340510.10.3604-3.56e-5
344531.2-0.8180-
350542.60.4070-3.51e-5
360557.00.3706-3.46e-5
370571.50.8766-3.41e-5
380585.80.8019-3.36e-5
390600.40.6258-3.31e-5
400614.60.8164-3.25e-5
410628.90.8534-3.20e-5
420643.10.1929-3.15e-5
430657.50.6785-3.10e-5
440671.80.5623-3.05e-5
450686.20.7090-3.00e-5
460700.50.5157-2.95e-5
470714.80.0765-2.89e-5
480729.20.4413-2.84e-5
490743.60.2660-2.79e-5
500758.00.2874-2.74e-5
510772.41.0218-2.69e-5
520786.70.6218-2.64e-5
530801.10.7205-2.58e-5
540815.41.2731-2.53e-5
550829.80.7234-2.48e-5
560844.10.8504-2.43e-5
570858.50.5348-2.38e-5
580872.70.2974-2.33e-5
590887.20.2326-2.28e-5
600901.51.0388-2.22e-5
610915.90.2176-2.17e-5
620930.30.1636-2.12e-5
630944.60.7971-2.07e-5
640959.10.6642-2.02e-5
650973.30.5806-1.97e-5
660987.60.3757-1.92e-5
6701002.00.3868-1.86e-5
6801016.30.5389-1.81e-5
6881043.2-0.8427-
6901048.30.3520-1.76e-5
7001062.50.5634-1.71e-5
7101076.80.4575-1.66e-5
7201091.10.3598-1.61e-5
7301105.50.5204-1.56e-5
7401119.90.2292-1.50e-5
7501134.30.2375-1.45e-5
7601148.80.9570-1.40e-5
7701163.10.5869-1.35e-5
7801177.40.2324-1.30e-5
7901191.70.3805-1.25e-5
8001206.00.1558-1.19e-5
8101220.50.1812-1.14e-5
8201234.80.2480-1.09e-5
8301249.00.2593-1.04e-5
8401263.30.1775-9.89e-6
8501277.70.2547-9.37e-6
8601291.80.1383-8.86e-6
8701306.10.4216-8.34e-6
8801320.50.2400-7.83e-6
8901334.80.3615-7.31e-6
9001349.30.6930-6.80e-6
9101363.60.1802-6.28e-6
9201377.90.5840-5.77e-6
9301392.20.2780-5.25e-6
9401406.50.2047-4.74e-6
9501420.70.3306-4.22e-6
9601434.90.0977-3.71e-6
9701449.20.4782-3.19e-6
9801463.50.1337-2.68e-6

GPU readings

106 rows as a table
StepGPUUse %Power W, whole chip, CPU and GPUMemory GB, whole GPUTemp C, one board zone
1002083.916.791.8
2000120.717.289.8
400100117.717.289.8
6000119.917.290.8
900100118.817.290.8
1100100114.220.490.8
13000121.320.491.8
1500100119.320.491.8
18000119.820.491.8
2000100116.523.791.8
22000120.323.791.8
2400100118.223.791.8
2700100114.423.791.8
29000120.623.791.8
3100100115.723.791.8
33000121.423.791.8
3440100122.723.796.8
3700100113.623.790.8
3900100122.623.790.8
4100100116.123.790.8
44000118.423.790.8
4600100116.923.790.8
48000118.423.790.8
5000100117.423.790.8
53000120.123.790.8
5500100117.723.790.8
5700100113.523.790.8
5900100119.023.791.8
62000114.423.791.8
6400100119.223.791.8
6600100114.423.791.8
680054123.123.791.8
6900100113.523.792.8
72000123.523.790.8
7400100113.523.790.8
7600100122.223.790.8
7800100114.623.790.8
81000122.723.790.8
8300100116.523.790.8
8500100113.823.790.8
8700100122.923.790.8
9000100114.323.790.8
92000121.323.790.8
9400100118.323.790.8
9700100114.923.790.8
99000120.823.790.8
10100100116.223.790.8
1032092118.02.192.8

Worked out

ReadoutValueWorked out from
Run complete100%1032 steps over 1032
Step time1.45 s1468.1 s over 1010 steps, between readings. Startup and intervals that hold an eval are left out
Steps per minute41.360 s over the step time of 1.45 s
Epoch interval497.6 sMean of 2 gaps between epoch marks
Train loss since first reading−1.30760.4733 at step 1032 minus 1.7809 at step 10
Lowest train loss0.0441Lowest of 104 readings, at step 1030
Eval accuracy, as shown84.7%0.8472 times 100, to one decimal. Changes on that plate are percentage points
Eval accuracy since first eval+0.02920.8472 minus 0.8180
Eval change+0.0045Eval 3 at 0.8472 minus eval 2 at 0.8427
Wrong answers68445 eval items minus 377 correct. 377 over 445 is the only count that gives 0.8472
Rank3 of 6By last eval accuracy, across this project's 6 runs with an eval
Best so far0.9580Highest last eval accuracy, run 1
Against the run before−0.04270.8472 minus 0.8899 on run 3
Lowest last train loss0.0241Lowest last train loss across this project's 6 runs, run 1
Items per second11.01 step over 1.45 s, times batch size 16
Train loss against run 3+0.23680.4733 minus 0.2366 on run 3, both at step 1032
Learning rate since first reading−8.20e-60.00e+0 minus 8.20e-6
Peak memory per GPU, whole GPU23.7 GBHighest sensor reading of any GPU, GPU 0, counting every process on it