CITADEL
StandbyNothing is live right now.
Eval accuracylast measured Wed 18 Nov, 08:43 UTC
0%25%50%75%100%0%25%50%75%100%0%25%50%75%100%

Familiar decider: Laya (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, 07:17 UTC
ran for
86:08
config
batch 8, seed 2 of 3

run3/3step1444/1446epoch2/3seed2/3batch8modellaya 421Mrev7b928d8items7701+675

Loss train 0.1202 ▼ −1.2951 eval 0.3603 lowest -0.6995 at step 1289 run before, run 2 ▲ +0.1880 step 1444 Laya's objective: soft cross-entropy plus a policy-gradient term, so it can go below zero.
Timebase
Range
Smoothing
Reference
288 readings, 30 min/div30 min/div10 min/div
-1012300:0030:0060:00phase starting, 0.0 s to 17.9 sphase training, 17.9 s to 86:08epoch 1, 28:42epoch 2, 57:25epoch 3, 86:08trainingepoch 1epoch 21.4153Lowest last train loss across this project's 3 runs, run 2lowest last -0.0678, run 21446 steps0.1202-1012300:0030:0060:00phase starting, 0.0 s to 17.9 sphase training, 17.9 s to 86:08epoch 1, 28:42epoch 2, 57:25epoch 3, 86:08trainingepoch 1epoch 21.4153Lowest last train loss across this project's 3 runs, run 2lowest last -0.0678, run 2eval 0.36031446 steps0.1202▼ −1.2951-1012300:0010:0020:0030:0040:0050:0060:0070:0080:00phase starting, 0.0 s to 17.9 sphase training, 17.9 s to 86:08epoch 1, 28:42epoch 2, 57:25epoch 3, 86:08trainingepoch 1epoch 21.4153Lowest last train loss across this project's 3 runs, run 2lowest last -0.0678, run 2eval 0.36031446 stepsrun 20.1202▼ −1.2951

Eval measures

Eval losslower is better
0.3603
02 →
ECEcalibration error, lower is better
0.0437
01

Elapsed

1722.8 s

epoch interval

1722.8 sto epoch 11722.8 sto epoch 21722.8 sto epoch 3

1722.8 s to epoch 1, 1722.8 s to epoch 2, 1722.8 s to epoch 3

1444 of 1446 steps, 2 of 3 epochs, 99.9%

3.58 s per step, 16.8 steps a minute

Against the other runs

1 of 3

by last eval accuracy

best so far 0.8622, run 3

▲ +0.0192 against run 2

Train minus eval

−0.2401

step 1444

train 0.1202, eval 0.3603

Set-up

7701 train, 675 eval

synthetic, batch size 8, grad accum 2, epochs 3, loss rlcd, seed 2

Seed spread, observed

0.0192
0.800.90

3 finished seeds

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 newer ▸

3 of 3 finished at step 1444 of 1446, 0.8622, config 1 of 1 (batch 8), seed 2 of 3, Wed 18 Nov, 07:17 UTC

▸
Runs as a table
StartedStatusEval accuracyStepsConfig
Wed 18 Nov, 07:17 UTC| finished0.86221444 of 1446batch 8, seed 2 of 3
Wed 18 Nov, 05:51 UTC| finished0.84301444 of 1446batch 8, seed 1 of 3
Wed 18 Nov, 04:24 UTC| finished0.85931444 of 1446batch 8, seed 0 of 3

Plate

Model
convaiinnovations/laya, 421M parameters revision 7b928d8
Data
Familiar decisions, shared draw, synthetic7701 train675 eval
Config
batch size 8, grad accum 2, epochs 3, loss rlcd, seed 2
Hardware
Radeon 8060S, ROCm

Log

Since startMark
0.0 sphase starting
17.9 sphase training
1722.8 sepoch 1
3445.6 sepoch 2
5168.4 sepoch 3
5168.4 seval 1
5 h 35 min laterwent public

Seeds

SeedStatusEval accuracy
seed 0| finished0.8593
seed 1| finished0.8430
seed 2| finished0.8622

observed spread 0.0192, 3 finished seeds

Readings

288 rows as a table

rows 1 to 100 of 288

StepSecondsTrain lossEval lossEval accuracyECE
517.91.4153---
1035.91.3986---
1553.80.2346---
2071.80.4170---
2589.70.5383---
30107.70.4620---
35125.60.4621---
40143.60.8409---
45161.51.6991---
50179.51.2447---
55197.41.7672---
60215.31.1995---
65233.31.0439---
70251.20.8982---
75269.20.6001---
80287.10.8157---
85305.10.9864---
90323.00.8205---
95341.01.2495---
100358.9-0.1212---
105376.90.0546---
110394.80.9118---
115412.81.2965---
120430.70.3179---
125448.60.9622---
130466.60.8906---
135484.50.5199---
140502.50.2978---
145520.40.5428---
150538.40.3071---
155556.31.1114---
160574.30.3121---
165592.20.8642---
170610.20.2324---
175628.10.2379---
180646.00.1593---
185664.00.5167---
190681.92.0904---
195699.90.7474---
200717.80.3211---
205735.81.0631---
210753.7-0.2235---
215771.70.0151---
220789.60.7841---
225807.60.2691---
230825.50.4510---
235843.50.5567---
240861.4-0.2332---
245879.30.6324---
250897.31.0250---
255915.2-0.6008---
260933.20.5805---
265951.10.5130---
270969.11.3178---
275987.01.0625---
2801005.00.3039---
2851022.90.3786---
2901040.90.4134---
2951058.81.3286---
3001076.80.6951---
3051094.70.7499---
3101112.60.6095---
3151130.61.2490---
3201148.51.4313---
3251166.50.5725---
3301184.40.0452---
3351202.40.1981---
3401220.30.4072---
3451238.30.4576---
3501256.2-0.1002---
3551274.2-0.0736---
3601292.10.2458---
3651310.00.3369---
3701328.00.9835---
3751345.90.5959---
3801363.90.7897---
3851381.80.4752---
3901399.80.5058---
3951417.70.3615---
4001435.70.8175---
4051453.60.6360---
4101471.60.3847---
4151489.5-0.0362---
4201507.50.7429---
4251525.41.0326---
4301543.3-0.0341---
4351561.30.3027---
4401579.20.8383---
4451597.20.7042---
4501615.10.2672---
4551633.1-0.3562---
4601651.00.3793---
4651669.00.4333---
4701686.90.3056---
4751704.90.5705---
4801722.80.2458---
4871740.70.9404---
4921758.7-0.0028---
4971776.60.8880---
5021794.60.5070---

Worked out

ReadoutValueWorked out from
Run complete99.9%1444 steps over 1446
Step time3.58 s5132.6 s over 1434 steps, between readings. Startup and intervals that hold an eval are left out
Steps per minute16.860 s over the step time of 3.58 s
Epoch interval1722.8 sMean of 2 gaps between epoch marks
Train loss since first reading−1.29510.1202 at step 1444 minus 1.4153 at step 5
Lowest train loss-0.6995Lowest of 288 readings, at step 1289
Eval accuracy, as shown86.2%0.8622 times 100, to one decimal. Changes on that plate are percentage points
Train minus eval loss−0.24010.1202 minus 0.3603, both at step 1444
Wrong answers93675 eval items minus 582 correct. 582 over 675 is the only count that gives 0.8622
Rank1 of 3By last eval accuracy, across this project's 3 runs with an eval
Best so far0.8622Highest last eval accuracy, run 3
Against the run before+0.01920.8622 minus 0.8430 on run 2
Lowest last train loss-0.0678Lowest last train loss across this project's 3 runs, run 2
Seed spread, observed0.0192High 0.8622 minus low 0.8430 across 3 finished seeds
Items per second2.21 step over 3.58 s, times batch size 8
Train loss against run 2+0.18800.1202 minus -0.0678 on run 2, both at step 1444