novels_all_20260225161648_5t6xstalenovels7.21M params20m 6s elapsed · Updated 47d ago
6L / 288D / 6H · helios · bpe · adamw· Created Feb 25, 2026 4:16 PM
Step 708 / 50,0001.4%
7.2347
Loss?
6.9177
Best Loss?
-5.5% from start
7.6543
Val Loss?
best: 7.6543
3.00e-4
Learning Rate?
3,824
Throughput?
tok/s (avg)
1,654
Speed?
ms/iter (avg)
0.565
Grad Norm?
avg: 0.619
3.61M
Tokens
processed
128ms
Forward
8% of step
1449ms
Backward
88% of step
20ms
GPU Sync
1% of step
568
GPU Ops
per step
0.4%
MFU
model FLOPS util
11.3x
Bwd/Fwd
ratio
Loss Curve ? click any chart to add markers
?
?
?
?
Architecture
Layers?6
Embedding?288
Heads?6
Vocab?2,000
Context?256
Dropout?0
Parameters?7.21M
Training Config
Total iters?50,000
Batch size?20
Max LR?0.0003
Optimizer?adamw
Backend?helios
Tokenizer?bpe
Seed?42
Weight decay?0.1
Grad clip?5
Eval interval?100
Throughput (tok/s)
Step Time (ms/iter)
GPU & VRAM
Perplexity
Train/Val Gap
Learning Rate
Grad Norm
Smoothed Loss (EMA)
Loss Velocity
Gradient Clipping
GPU Operations
Step Time Breakdown
Forward
Backward
Grad Norm
Optimizer
GPU Sync
Data
Timing Phase Lines
Backward / Forward Ratio
Evolutionary Analysis (Symbiogenesis)
1.77
Wt Entropy
bits
20.0
Eff. Rank
7.0108
Free Energy
3.892
Pop Entropy
nats
0.0786
Complexity
0.0465
Fitness
695
CUSUM
alerts
12
Batch Size
adaptive
CUSUM Statistical Monitors
Information Bottleneck (MI)
MI Analysis Pending
Checkpoints (0) ?
No checkpoints saved
Sample Generations (3)
#CheckpointPrompt (preview)Generated
1-The 47d ago
Prompt
The
Output
The something thought contwasforERations intelligcould decreferimple people en the s to le n'trequbeforGPheiter couldyouof sembctustill ely sput . They CH
abclaude parresumcationed in fatheme t as a betweenplac. And containight ight
2-Once upon a time47d ago
Prompt
Once upon a time
Output
Once upon a timebre there promptdescriTraves ativfirste. The that was ese they were pter eathot on ction- storFetchastwas slolininstESode ================================otharnessmy, vers she end ken theint Lisa had managch particularension that would gupetos. s were CHAPTH
3-He walked into47d ago
Prompt
He walked into
Output
He walked into callpurOptionrun turnhaddro.
It anc," ing that lessum thanust ed.
anddata place ponentexact. Nboth .
It ind Cdatptionfindponweeous because. The thwhen think fromday impl someconversationeven thering. sid cre==gerdict
Model Config (JSON)
{
"vocabSize": 2000,
"blockSize": 256,
"nLayer": 6,
"nEmbd": 288,
"nHead": 6,
"dropout": 0,
"ffnActivation": "swiglu",
"ffnDim": 768
}Training Config (JSON)
{
"iters": 50000,
"batchSize": 20,
"lr": 0.0003,
"lrMin": 0,
"warmupIters": 500,
"beta1": 0.9,
"beta2": 0.95,
"eps": 1e-8,
"weightDecay": 0.1,
"gradClip": 5,
"evalInterval": 100,
"evalIters": 10,
"seed": 42,
"backend": "helios",
"tokenizer": "bpe",
"optimizer": "adamw",
"logLevel": "info",
"trace": false,
"gradAccumSteps": 1,
"sampleInterval": 100,
"spikeThreshold": 10,
"syncEvery": 1,
"gcEvery": 0,
"packed": false,
"symbio": true,
"symbioConfig": {
"cusumSensitivity": 4,
"cusumBaselineWindow": 5,
"metricsInterval": 10,
"trackWeightEntropy": true,
"trackEffectiveRank": true,
"trackFreeEnergy": true,
"trackMIProfiles": false,
"trackPopulationMetrics": true,
"freeEnergyBeta": 0.01,
"miNumBins": 30,
"adaptiveBatch": true,
"batchMin": 8,
"batchMax": 64,
"batchStep": 4,
"calmStepsBeforeRestore": 200,
"fitnessAlpha": 1,
"complexityMode": "entropy",
"diversityBonus": 0.1,
"diversityDecay": "cosine",
"searchMode": "composed-activation-search",
"activationPool": [
"gelu",
"relu",
"silu",
"swiglu",
"universal",
"kan_spline"
],
"searchStrategy": "evolutionary",
"populationSize": 8,
"generations": 250,
"selectionStrategy": "topk",
"tournamentK": 3,
"mutationRate": 0.7,
"stepsPerCandidate": 25,
"rankBy": "valLoss",
"perfWeight": 0,
"stabilityWeight": 0,
"writeReport": true,
"writeCandidates": true,
"writeSummary": true,
"basisPool": [
"silu",
"relu",
"gelu",
"identity",
"square"
],
"maxGraphDepth": 4,
"maxGraphNodes": 10
}
}