20260223122645_g850completedchat9.99M params43m 51s elapsed · Updated 36d ago
6L / 256D / 8H · helios · bpe-4k · adamw· Created Feb 23, 2026 12:27 PM
Step 5,000 / 5,000100.0%
4.1786
Loss?
3.9733
Best Loss?
-49.9% from start
4.7442
Val Loss?
best: 4.7442
5.00e-6
Learning Rate?
7,733
Throughput?
tok/s (avg)
530
Speed?
ms/iter (avg)
65.845
Grad Norm?
avg: 57.009
20.40M
Tokens
processed
61ms
Forward
11% of step
410ms
Backward
77% of step
24ms
GPU Sync
5% of step
1,064
GPU Ops
per step
1.5%
MFU
model FLOPS util
6.8x
Bwd/Fwd
ratio
Loss Curve ? click any chart to add markers
?
?
?
?
Architecture
Layers?6
Embedding?256
Heads?8
Vocab?4,000
Context?256
Dropout?0
Parameters?9.99M
Training Config
Total iters?5,000
Batch size?16
Max LR?0.00005
Optimizer?adamw
Backend?helios
Tokenizer?bpe-4k
Seed?42
Weight decay?0.1
Grad clip?5
Eval interval?500
Throughput (tok/s)
Step Time (ms/iter)
GPU & VRAM
Perplexity
Train/Val Gap
Learning Rate
Grad Norm
Smoothed Loss (EMA)
Loss Velocity
Gradient Clipping
No clipping data
GPU Operations
Step Time Breakdown
Forward
Backward
Grad Norm
Optimizer
GPU Sync
Data
Timing Phase Lines
Backward / Forward Ratio
Sample Generations (5)
#CheckpointPrompt (preview)Generated
1-<|user|> Hello, how are you? <|assistant|>49d ago
Prompt
<|user|> Hello, how are you? <|assistant|>
Output
<|user|> Hello, how are you? <|assistant|>
<|assistant|> i ctext|>
<|assistant|> ygpfit <|user|> Ied <|user|> w<|end_of_t <|assistant|> in <|assistant|> <|assistant|> I s and et|>
<|assistant|> I a <|user|> Tos the dw
2-<|user|> What do you like to do for fun? <|assistant|>49d ago
Prompt
<|user|> What do you like to do for fun? <|assistant|>
Output
<|user|> What do you like to do for fun? <|assistant|>d a fe<|end_of_t|>
<|assistant|> MS of a cts s<|end_, it cd_of_textext|>
<|assistant|> I <|user|> A<|assistant|> I me<|end
3-<|user|> Tell me about yourself. <|assistant|>49d ago
Prompt
<|user|> Tell me about yourself. <|assistant|>
Output
<|user|> Tell me about yourself. <|assistant|>
<|user|> i drerwocha<|assistant|> Ta <|user|> G <|assistant|> I and your a <|user|> i <|assistant|> kreced <|assistant|> Ghya <|user|> M<|user|> in <|user|> i <|assistant|> in <|assistant|> O <|assistant|> I I the e<|user|> a
4-In the beginning 49d ago
Prompt
In the beginning
Output
In the beginning and the ender, destass, Combenea be rect to gon<|end_of_text|>
<|assistant|> i got my ptoof <|user|> I did for it but i know thats <|assistant|> I was just rechai clearo <|user|> i just didont get it <|user|> Waaint whahatswise <|assistant|> okay <|assistant|> AL AD is <|assistant|> Wim i like i
5-We the People of 49d ago
Prompt
We the People of
Output
We the People of some currenna sion that darmal <|user|> Sh, redyyes, your eud <|user|> Yun is better (i don't get itchu sea joo <|assistant|> I d<|end_of_text|>
<|user|> Ohmmm just s? <|assistant|> Im <|assistant|> NE <|assistant|> HAOFE MIN <|assistant|> ACRE DS SS G T
Model Config (JSON)
{
"vocabSize": 4000,
"blockSize": 256,
"nLayer": 6,
"nEmbd": 256,
"nHead": 8,
"dropout": 0
}Training Config (JSON)
{
"iters": 5000,
"batchSize": 16,
"lr": 0.00005,
"lrMin": 0.000005,
"warmupIters": 500,
"beta1": 0.9,
"beta2": 0.95,
"eps": 0.000001,
"weightDecay": 0.1,
"gradClip": 5,
"evalInterval": 500,
"evalIters": 10,
"seed": 42,
"backend": "helios",
"tokenizer": "bpe-4k",
"optimizer": "adamw",
"logLevel": "info",
"trace": true,
"gradAccumSteps": 1,
"sampleInterval": 500
}