CITADEL
StandbyNothing is live right now.
Eval accuracylast measured Wed 18 Nov, 02:22 UTC
%
0.8480424 of 500 eval, 76 wrong
▲ +83.60 pts across 4 evals▲ +0.20 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 (backfilled)

finishedRadeon 8060S

One of the small models that decide whether Familiar, a desktop companion, reacts to something happening on the PC, and how. 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
Wed 18 Nov, 01:40 UTC
ran for
31:20
config
learning rate 0.00005, batch 16, seed 0

run3/5step1125/1125epoch3/3config3/3seed0lr0.00005batch16modellaya 395Mrev7b928d8items6000+500

Eval measures

Elapsed

626.8 s

epoch interval

626.8 sto eval 2626.8 sto eval 3626.8 sto eval 4

626.8 s to eval 2, 626.8 s to eval 3, 626.8 s to eval 4

1125 of 1125 steps, 3 of 3 epochs, 100%

1.67 s per step, 35.9 steps a minute

Eval accuracy by eval

+0.0020

▲ on the eval before

eval 1 0.0120, eval 4 0.8480

Against the other runs

5 of 5

by last eval accuracy

best so far 0.9100, run 2

▼ −0.0620 against run 2

Set-up

6000 train, 500 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

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, 0.8480, config 3 of 3 (learning rate 0.00005, batch 16), seed 0, Wed 18 Nov, 01:40 UTC

Runs as a table
StartedStatusEval accuracyStepsConfig
Wed 18 Nov, 03:42 UTC| finished0.89401125 of 1125learning rate 0.00002, batch 16, seed 2 of 3
Wed 18 Nov, 03:00 UTC| finished0.87601125 of 1125learning rate 0.00002, batch 16, seed 1 of 3
Wed 18 Nov, 01:40 UTC| finished0.84801125 of 1125learning rate 0.00005, batch 16, seed 0
Wed 18 Nov, 00:58 UTC| finished0.91001125 of 1125learning rate 0.00002, batch 16, seed 0 of 3
Wed 18 Nov, 00:16 UTC| finished0.85601125 of 1125learning rate 0.0001, batch 16, seed 0

Plate

Model
convaiinnovations/laya, 395M parameters revision 7b928d8
Data
Familiar decisions, shared draw, synthetic6000 train500 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
0.0 seval 1
626.8 seval 2
626.8 sepoch 1
1253.6 seval 3
1253.6 sepoch 2
1880.4 seval 4
1880.4 sepoch 3
11 h 12 min laterwent public

Readings

StepSecondsEval accuracy
00.00.0120
375626.80.8140
7501253.60.8460
11251880.40.8480

Worked out

ReadoutValueWorked out from
Run complete100%1125 steps over 1125
Step time1.67 s626.8 s between epoch marks, over 375 steps an epoch. The reading interval holds the eval, so it is left out
Steps per minute35.960 s over the step time of 1.67 s
Epoch interval626.8 sMean of 2 gaps between epoch marks
Eval accuracy, as shown84.8%0.8480 times 100, to one decimal. Changes on that plate are percentage points
Eval accuracy since first eval+0.83600.8480 minus 0.0120
Eval change+0.0020Eval 4 at 0.8480 minus eval 3 at 0.8460
Wrong answers76500 eval items minus 424 correct. 424 over 500 is the only count that gives 0.8480
Rank5 of 5By last eval accuracy, across this project's 5 runs with an eval
Best so far0.9100Highest last eval accuracy, run 2
Against the run before−0.06200.8480 minus 0.9100 on run 2
Items per second9.61 step over 1.67 s, times batch size 16