CITADEL
StandbyNothing is live right now.
Eval accuracylast measured Sat 21 Nov, 11:39 UTC
%
0.9580479 of 500 eval, 21 wrong
▲ +3.60 pts across 2 evals▲ +3.60 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, 08:25 UTC
ran for
184:45
config
batch 16, seed 0

run1/6step7274/7274epoch2/2config1/5seed0batch16modellaya 395Mrev7b928d8items58183+500

Loss train 0.0241 ▼ −2.1548 lowest 0.0007 at step 6970
Timebase
Range
Smoothing
728 readings, 60 min/div60 min/div30 min/div
012300:0060:00120:00180:00phase starting, 0.0 s to 56.5 sphase training, 56.5 s to 184:45epoch 1, 92:47epoch 2, 184:27trainingepoch 1−2.1548since first2.1789Lowest last train loss across this project's 6 runs, run 1lowest last 0.0241, run 10.0241012300:0060:00120:00180:00phase starting, 0.0 s to 56.5 sphase training, 56.5 s to 184:45epoch 1, 92:47epoch 2, 184:27trainingepoch 12.1789Lowest last train loss across this project's 6 runs, run 1lowest last 0.0241, run 10.0241▼ −2.1548012300:0030:0060:0090:00120:00150:00180:00phase starting, 0.0 s to 56.5 sphase training, 56.5 s to 184:45epoch 1, 92:47epoch 2, 184:27trainingepoch 12.1789Lowest last train loss across this project's 6 runs, run 1lowest last 0.0241, run 10.0241▼ −2.1548

Eval measures

Elapsed

5499.9 s

epoch interval

5561.0 sto eval 15506.2 sto epoch 2

5561.0 s to eval 1, 5506.2 s to epoch 2

7274 of 7274 steps, 2 of 2 epochs, 100%

1.52 s per step, 39.6 steps a minute

Eval accuracy by eval

+0.0360

▲ on the eval before

eval 1 0.9220, eval 2 0.9580

Against the other runs

1 of 6

by last eval accuracy

best so far 0.9580, run 1

Set-up

58183 train, 500 eval

synthetic, batch size 16, epochs 2, max length 512, precision bf16-mixed, loss ce, optimizer adamw, scheduler linear, seed 0

Learning rate

0.00e+0

▼ −4.59e-7 since first

On one axis

phase starting, 0.0 s to 56.5 sphase training, 56.5 s to 184:45E1E2traininglowest last 0.0241Train loss0.0241▼ −2.15483.000.002.17890.0241last ▼ −2.1548best 0.9580, run 1Last-value range, 6 runs, 0.5618 to 0.9580Eval accuracy0.9580▲ +0.03601.000.400.92200.9580last ▲ +0.0360Learning rate0.00e+0▼ −4.59e-72.0e-50.0e+04.59e-7last 0.00e+0Power, whole chip, CPU and GPU118 W▲ +3 W125 W110 W115 Wlast 118 WHottest, one board zone94 °C▲ +3 °C96 °C88 °C91 °Clast 94 °C00:0030:0060:0090:00120:00150:00180:00
GPUpower: whole chip, CPU and GPUtemperature: one board zone

GPU 0

Wwhole chip, CPU and GPU

118 W whole chip, CPU and GPU

use90%
memory, whole GPU2.1 GB

peak 42.9 GB, whole GPU

94 °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 ◂ older

1 of 6 finished, 0.9580, config 1 of 5 (batch 16), seed 0, Sat 21 Nov, 08:25 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, full pool, synthetic58183 train500 eval
Config
batch size 16, epochs 2, 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
10.6 swent public
56.5 sphase training
5561.0 seval 1
5567.3 sepoch 1
11067.2 sepoch 2
11085.4 seval 2

Readings

729 rows as a table

rows 1 to 100 of 729

StepSecondsTrain lossEval accuracyLearning rate
1056.52.1789-4.59e-7
2071.61.6822-9.17e-7
3086.31.3487-1.38e-6
40101.41.4927-1.83e-6
50116.31.4611-2.29e-6
60131.51.4892-2.75e-6
70146.82.1618-3.21e-6
80162.01.7026-3.67e-6
90177.21.5567-4.13e-6
100192.41.5055-4.59e-6
110207.51.8069-5.05e-6
120222.41.4036-5.50e-6
130237.50.8940-5.96e-6
140252.71.1897-6.42e-6
150267.81.0536-6.88e-6
160282.91.2229-7.34e-6
170297.91.7807-7.80e-6
180313.22.1427-8.26e-6
190328.31.6395-8.72e-6
200343.71.3869-9.17e-6
210358.71.5725-9.63e-6
220373.81.0638-1.01e-5
230388.91.1916-1.06e-5
240403.90.9795-1.10e-5
250418.91.3584-1.15e-5
260433.91.4357-1.19e-5
270448.90.9810-1.24e-5
280464.10.8464-1.28e-5
290479.21.0909-1.33e-5
300494.31.7121-1.38e-5
310509.40.7783-1.42e-5
320524.40.8582-1.47e-5
330539.61.2662-1.51e-5
340554.60.9215-1.56e-5
350569.80.3648-1.61e-5
360584.80.7577-1.65e-5
370599.90.7853-1.70e-5
380615.01.1487-1.74e-5
390630.00.4925-1.79e-5
400645.12.4242-1.83e-5
410660.10.7531-1.88e-5
420675.20.9126-1.93e-5
430690.21.1416-1.97e-5
440705.30.9243-2.00e-5
450720.70.4533-2.00e-5
460735.70.4354-1.99e-5
470750.90.2703-1.99e-5
480765.51.0258-1.99e-5
490780.80.7707-1.98e-5
500796.01.0971-1.98e-5
510811.40.4382-1.98e-5
520826.60.9089-1.98e-5
530841.70.3342-1.97e-5
540856.90.6696-1.97e-5
550872.20.8295-1.97e-5
560887.40.7242-1.96e-5
570902.50.8170-1.96e-5
580917.80.7820-1.96e-5
590933.00.7727-1.95e-5
600948.30.4671-1.95e-5
610963.30.7176-1.95e-5
620978.30.8050-1.95e-5
630993.40.4991-1.94e-5
6401008.70.1609-1.94e-5
6501023.80.6184-1.94e-5
6601038.80.3393-1.93e-5
6701053.90.7547-1.93e-5
6801069.00.0789-1.93e-5
6901084.20.5714-1.93e-5
7001099.20.4998-1.92e-5
7101114.20.4044-1.92e-5
7201129.60.4484-1.92e-5
7301144.91.0591-1.91e-5
7401160.10.5262-1.91e-5
7501175.00.9890-1.91e-5
7601190.20.3041-1.91e-5
7701205.10.6143-1.90e-5
7801220.31.1903-1.90e-5
7901235.40.5783-1.90e-5
8001251.00.4544-1.89e-5
8101266.30.2024-1.89e-5
8201281.50.7006-1.89e-5
8301296.60.4571-1.88e-5
8401311.80.4667-1.88e-5
8501326.70.5183-1.88e-5
8601341.80.8035-1.88e-5
8701356.90.7047-1.87e-5
8801371.90.3818-1.87e-5
8901387.10.4858-1.87e-5
9001402.70.5513-1.86e-5
9101417.60.3146-1.86e-5
9201432.60.3971-1.86e-5
9301447.90.4950-1.86e-5
9401462.90.3899-1.85e-5
9501478.00.9411-1.85e-5
9601493.20.3550-1.85e-5
9701508.30.3436-1.84e-5
9801523.50.6829-1.84e-5
9901538.90.6824-1.84e-5
10001554.30.6985-1.84e-5

GPU readings

729 rows as a table

rows 1 to 100 of 342

StepGPUUse %Power W, whole chip, CPU and GPUMemory GB, whole GPUTemp C, one board zone
100100115.117.390.8
2000120.820.688.8
400100117.123.989.8
6000121.231.590.8
9000115.631.591.8
1100100115.631.592.8
13000120.931.592.8
1500100113.731.592.8
1700100118.931.593.8
19000120.031.592.8
2100100115.931.592.8
23000120.431.593.8
2600100116.531.593.8
28000121.131.593.8
3000100115.431.593.8
32000122.931.593.8
3400100113.231.594.8
3600100119.731.593.8
3900100113.431.593.8
4100100119.131.593.8
4300100113.031.593.8
4500100116.431.593.8
47000116.331.593.8
4900100115.731.593.8
5100100119.231.593.8
53000117.831.593.8
5600100112.531.592.8
5800100116.631.592.8
60000121.831.593.8
6200100115.231.592.8
6400100122.631.592.8
6600100112.831.592.8
6800100116.831.593.8
71000119.531.592.8
73000115.131.592.8
7500100115.831.592.8
77000118.531.592.8
7900100112.931.592.8
8100100113.931.592.8
8300100117.831.592.8
86000113.331.592.8
8800100119.131.591.8
9000100122.235.292.8
9200100114.235.292.8
9400100122.035.292.8
96000113.735.292.8
9800100114.735.292.8
10000100116.138.992.8
103000123.538.992.8
105000120.738.992.8
10700100115.938.992.8
109000119.938.992.8
11100100115.338.992.8
113000121.938.992.8
11500100112.538.992.8
11800100116.138.993.8
120000119.838.992.8
12200100115.838.993.8
124000121.238.993.8
12600100111.938.993.8
12800100115.738.993.8
130000121.538.993.8
13300100114.538.993.8
13500100118.338.993.8
137000113.038.993.8
13900100116.838.993.8
14100100122.138.993.8
143000114.438.993.8
14500100116.938.993.8
147000122.438.993.8
15000100115.538.993.8
152000123.338.993.8
154000116.838.993.8
15600100117.838.992.8
158000121.038.992.8
160000116.638.993.8
16200100115.938.993.8
165000119.838.993.8
16700100116.038.993.8
169000121.538.993.8
17100100113.338.993.8
17300100119.138.993.8
175000113.338.993.8
17700100118.838.993.8
180000113.838.993.8
18200100116.638.993.8
184000120.138.993.8
18600100114.438.993.8
18800100121.938.993.8
19000100112.238.992.8
19200100118.138.992.8
195000119.838.992.8
197000120.238.993.8
19900100116.838.993.8
201000122.538.993.8
20300100113.738.993.8
20500100116.138.993.8
207000119.438.992.8
21000100115.238.992.8
21200100122.338.992.8

Worked out

ReadoutValueWorked out from
Run complete100%7274 steps over 7274
Step time1.52 s10999.1 s over 7257 steps, between readings. Startup and intervals that hold an eval are left out
Steps per minute39.660 s over the step time of 1.52 s
Epoch interval5499.9 sThe one gap between epoch marks
Train loss since first reading−2.15480.0241 at step 7274 minus 2.1789 at step 10
Lowest train loss0.0007Lowest of 728 readings, at step 6970
Eval accuracy, as shown95.8%0.9580 times 100, to one decimal. Changes on that plate are percentage points
Eval accuracy since first eval+0.03600.9580 minus 0.9220
Eval change+0.0360Eval 2 at 0.9580 minus eval 1 at 0.9220
Wrong answers21500 eval items minus 479 correct. 479 over 500 is the only count that gives 0.9580
Rank1 of 6By last eval accuracy, across this project's 6 runs with an eval
Best so far0.9580Highest last eval accuracy, run 1
Lowest last train loss0.0241Lowest last train loss across this project's 6 runs, run 1
Items per second10.61 step over 1.52 s, times batch size 16
Learning rate since first reading−4.59e-70.00e+0 minus 4.59e-7
Peak memory per GPU, whole GPU42.9 GBHighest sensor reading of any GPU, GPU 0, counting every process on it