CITADEL
StandbyNothing is live right now.
Eval accuracylast measured Tue 3 Nov, 01:38 UTC
0%25%50%75%100%0%25%50%75%100%0%25%50%75%100%

Claviger

finishedRadeon 8060S

A small model that reads one Discord message and decides whether an agent should reply, and which one. Fine-tuned from Laya on invented chat.

started
Tue 3 Nov, 01:26 UTC
ran for
12:49
config
batch 8, seed 2 of 3

run14/14step150/150epoch5/5config6/6seed2/3batch8modellaya 421Mrev7b928d8items479+130

Loss train -0.0028 ▼ −0.8513 eval 0.6889 lowest -0.4034 at step 100 run before, run 13 ▼ −0.0028 step 150 Laya's objective: soft cross-entropy plus a policy-gradient term, so it can go below zero.
Timebase
Range
Smoothing
Reference
30 readings, 5 min/div5 min/div2 min/div
-202400:0005:0010:00phase starting, 0.0 s to 77.5 sphase training, 77.5 s to 12:42phase eval, 12:42 to 12:49epoch 2, 05:06epoch 3, 07:13epoch 4, 09:18epoch 5, 11:29trainingE2E3E4E50.8485Lowest last train loss across this project's 14 runs, run 1lowest last -0.0811, run 1-0.0028-202400:0005:0010:00phase starting, 0.0 s to 77.5 sphase training, 77.5 s to 12:42phase eval, 12:42 to 12:49epoch 1, 02:57epoch 2, 05:06epoch 3, 07:13epoch 4, 09:18epoch 5, 11:29trainingepoch 1epoch 2epoch 3epoch 4epoch 50.8485Lowest last train loss across this project's 14 runs, run 1lowest last -0.0811, run 1eval 0.6889-0.0028▼ −0.8513-202400:0002:0004:0006:0008:0010:0012:00phase starting, 0.0 s to 77.5 sphase training, 77.5 s to 12:42phase eval, 12:42 to 12:49epoch 1, 02:57epoch 2, 05:06epoch 3, 07:13epoch 4, 09:18epoch 5, 11:29startingtrainingepoch 1epoch 2epoch 3epoch 4epoch 50.8485Lowest last train loss across this project's 14 runs, run 1lowest last -0.0811, run 1eval 0.6889-0.0028▼ −0.8513

Eval measures

Eval losslower is better
0.6889
02 →
ECEcalibration error, lower is better
0.1509
01
Brierlower is better
0.3399
02 →

Elapsed

128.1 s

epoch interval

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

4.22 s per step, 14.2 steps a minute

Against the other runs

11 of 14

by last eval accuracy

best so far 0.8385, run 5

▼ −0.0231 against run 13

Train minus eval

−0.6917

step 150

train -0.0028, eval 0.6889

Set-up

479 train, 130 eval

synthetic, batch size 8, grad accum 2, epochs 5, loss rlcd, seed 2

Seed spread, observed

0.0231
0.800.85

3 finished seeds

Samples per second

4.13

▼ −0.07 since first

Learning rate

1.00e-6

▼ −2.40e-5 since first

GPU memory, this run

6.5 GB

On one axis

phase starting, 0.0 s to 77.5 sphase training, 77.5 s to 12:42phase eval, 12:42 to 12:49E2E3E4E5traininglowest last -0.0811Train loss-0.0028▼ −0.85134.00-2.000.8485-0.0028last ▼ −0.8513Samples per second4.13▼ −0.07524.20Learning rate1.00e-6▼ −2.40e-53.0e-50.0e+02.50e-5Power, whole chip, CPU and GPU84 W▲ +0 W85 W83 W83 W84 Wlast ▲ +0 WHottest, one board zone94 °C▼ −1 °C95 °C94 °C95 °Clast 94 °C00:0002:0004:0006:0008:0010:0012:00
GPUpower: whole chip, CPU and GPUtemperature: one board zone

GPU 0

Wwhole chip, CPU and GPU

84 W whole chip, CPU and GPU

use0%
memory, whole GPU13.3 GB

this run's peak 8.0 GB

94 °C one board zone

history 0 to 84 W

Input

  • Runs14 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.

Runsheight is eval accuracyall recordings ▸ | finished× failed■ stopped○ lost● live newer ▸
0.850.750.850.75finished, eval accuracy 0.8077, config 1 of 6 (batch 8), seed 2, Mon 2 Nov, 13:19 UTCfinished, eval accuracy 0.8231, config 2 of 6 (batch 8), seed 2 of 3, Mon 2 Nov, 14:03 UTCfinished, eval accuracy 0.8231, config 3 of 6 (batch 8), seed 2, Mon 2 Nov, 14:09 UTCfinished, eval accuracy 0.7846, config 2 of 6 (batch 8), seed 0 of 3, Mon 2 Nov, 14:16 UTCfinished, eval accuracy 0.8385, config 2 of 6 (batch 8), seed 1 of 3, Mon 2 Nov, 14:22 UTCfinished, eval accuracy 0.7923, config 4 of 6 (batch 8), seed 0 of 3, Mon 2 Nov, 18:15 UTCfinished, eval accuracy 0.8000, config 4 of 6 (batch 8), seed 1 of 3, Mon 2 Nov, 18:22 UTCfinished, eval accuracy 0.8154, config 4 of 6 (batch 8), seed 2 of 3, Mon 2 Nov, 18:28 UTCfinished, eval accuracy 0.7923, config 5 of 6 (batch 8), seed 0 of 3, Mon 2 Nov, 18:46 UTCfinished, eval accuracy 0.8385, config 5 of 6 (batch 8), seed 1 of 3, Mon 2 Nov, 18:52 UTCfinished, eval accuracy 0.8000, config 5 of 6 (batch 8), seed 2 of 3, Mon 2 Nov, 18:59 UTCfinished, eval accuracy 0.8231, config 6 of 6 (batch 8), seed 0 of 3, Tue 3 Nov, 00:59 UTCfinished, eval accuracy 0.8231, config 6 of 6 (batch 8), seed 1 of 3, Tue 3 Nov, 01:13 UTCfinished, eval accuracy 0.8000, config 6 of 6 (batch 8), seed 2 of 3, Tue 3 Nov, 01:26 UTC0.850.75finished, eval accuracy 0.8077, config 1 of 6 (batch 8), seed 2, Mon 2 Nov, 13:19 UTCfinished, eval accuracy 0.8231, config 2 of 6 (batch 8), seed 2 of 3, Mon 2 Nov, 14:03 UTCfinished, eval accuracy 0.8231, config 3 of 6 (batch 8), seed 2, Mon 2 Nov, 14:09 UTCfinished, eval accuracy 0.7846, config 2 of 6 (batch 8), seed 0 of 3, Mon 2 Nov, 14:16 UTCfinished, eval accuracy 0.8385, config 2 of 6 (batch 8), seed 1 of 3, Mon 2 Nov, 14:22 UTCfinished, eval accuracy 0.7923, config 4 of 6 (batch 8), seed 0 of 3, Mon 2 Nov, 18:15 UTCfinished, eval accuracy 0.8000, config 4 of 6 (batch 8), seed 1 of 3, Mon 2 Nov, 18:22 UTCfinished, eval accuracy 0.8154, config 4 of 6 (batch 8), seed 2 of 3, Mon 2 Nov, 18:28 UTCfinished, eval accuracy 0.7923, config 5 of 6 (batch 8), seed 0 of 3, Mon 2 Nov, 18:46 UTCfinished, eval accuracy 0.8385, config 5 of 6 (batch 8), seed 1 of 3, Mon 2 Nov, 18:52 UTCfinished, eval accuracy 0.8000, config 5 of 6 (batch 8), seed 2 of 3, Mon 2 Nov, 18:59 UTCfinished, eval accuracy 0.8231, config 6 of 6 (batch 8), seed 0 of 3, Tue 3 Nov, 00:59 UTCfinished, eval accuracy 0.8231, config 6 of 6 (batch 8), seed 1 of 3, Tue 3 Nov, 01:13 UTCfinished, eval accuracy 0.8000, config 6 of 6 (batch 8), seed 2 of 3, Tue 3 Nov, 01:26 UTC

14 of 14 finished, 0.8000, config 6 of 6 (batch 8), seed 2 of 3, Tue 3 Nov, 01:26 UTC

▸
Runs as a table
StartedStatusEval accuracyStepsConfig
Tue 3 Nov, 01:26 UTC| finished0.8000150 of 150batch 8, seed 2 of 3
Tue 3 Nov, 01:13 UTC| finished0.8231150 of 150batch 8, seed 1 of 3
Tue 3 Nov, 00:59 UTC| finished0.8231150 of 150batch 8, seed 0 of 3
Mon 2 Nov, 18:59 UTC| finished0.8000150 of 150batch 8, seed 2 of 3
Mon 2 Nov, 18:52 UTC| finished0.8385150 of 150batch 8, seed 1 of 3
Mon 2 Nov, 18:46 UTC| finished0.7923150 of 150batch 8, seed 0 of 3
Mon 2 Nov, 18:28 UTC| finished0.8154150 of 150batch 8, seed 2 of 3
Mon 2 Nov, 18:22 UTC| finished0.8000150 of 150batch 8, seed 1 of 3
Mon 2 Nov, 18:15 UTC| finished0.7923150 of 150batch 8, seed 0 of 3
Mon 2 Nov, 14:22 UTC| finished0.8385150 of 150batch 8, seed 1 of 3
Mon 2 Nov, 14:16 UTC| finished0.7846150 of 150batch 8, seed 0 of 3
Mon 2 Nov, 14:09 UTC| finished0.8231150 of 150batch 8, seed 2
Mon 2 Nov, 14:03 UTC| finished0.8231150 of 150batch 8, seed 2 of 3
Mon 2 Nov, 13:19 UTC| finished0.8077150 of 150batch 8, seed 2

Plate

Model
convaiinnovations/laya, 421M parameters revision 7b928d8
Data
Synthetic Discord triage set, synthetic479 train130 eval
Config
batch size 8, grad accum 2, epochs 5, loss rlcd, seed 2
Hardware
Radeon 8060S, ROCm

Log

Since startMark
0.0 sphase starting
6.8 swent public
77.5 sphase training
177.6 sepoch 1
306.8 sepoch 2
434.0 sepoch 3
558.1 sepoch 4
690.0 sepoch 5
762.9 sphase eval
762.9 seval 1

Worked out

ReadoutValueWorked out from
Run complete100%150 steps over 150
Step time4.22 s612.5 s over 145 steps, between readings. Startup and intervals that hold an eval are left out
Steps per minute14.260 s over the step time of 4.22 s
Epoch interval128.1 sMean of 4 gaps between epoch marks
Train loss since first reading−0.8513-0.0028 at step 150 minus 0.8485 at step 5
Lowest train loss-0.4034Lowest of 30 readings, at step 100
Eval accuracy, as shown80.0%0.8000 times 100, to one decimal. Changes on that plate are percentage points
Train minus eval loss−0.6917-0.0028 minus 0.6889, both at step 150
Wrong answers26130 eval items minus 104 correct. 104 over 130 is the only count that gives 0.8000
Rank11 of 14By last eval accuracy, across this project's 14 runs with an eval
Best so far0.8385Highest last eval accuracy, run 5
Against the run before−0.02310.8000 minus 0.8231 on run 13
Lowest last train loss-0.0811Lowest last train loss across this project's 14 runs, run 1
Seed spread, observed0.0231High 0.8231 minus low 0.8000 across 3 finished seeds
Train loss against run 13−0.0028-0.0028 minus 0.0000 on run 13, both at step 150
Samples per second since first reading−0.074.13 minus 4.20
Learning rate since first reading−2.40e-51.00e-6 minus 2.50e-5
This run's peak memory per GPU8.0 GBAs the run reports it: its own memory, not the whole GPU's