chat_clean_20260225055124_w4hwcompletedunknown425.3K params10s elapsed · Updated 9m ago
2L / 64D / 2H · cpu_ref · bpe · adamw· Created Feb 25, 2026 5:53 AM
Step 20 / 20100.0%
7.4278
Loss?
7.3990
Best Loss?
-2.5% from start
7.4137
Val Loss?
best: 7.4137
3.26e-5
Learning Rate?
249
Throughput?
tok/s (avg)
520
Speed?
ms/iter (avg)
1.472
Grad Norm?
avg: 1.533
2.6K
Tokens
processed
160ms
Forward
31% of step
321ms
Backward
62% of step
0ms
GPU Sync
0% of step
0
GPU Ops
per step
0.0%
MFU
model FLOPS util
2.0x
Bwd/Fwd
ratio
Loss Curve ?
Validation / selection
Run diagnostics
Gradient + Loss
Train/val loss with validation, warmup, overfit, and instability markers.
Search-aware view
Architecture
Layers?2
Embedding?64
Heads?2
Vocab?2,000
Context?64
Dropout?0
Parameters?425.3K
Training Config
Total iters?20
Batch size?2
Max LR?0.0003
Optimizer?adamw
Backend?cpu_ref
Tokenizer?bpe
Seed?42
Weight decay?0.1
Grad clip?1
Eval interval?10
GPU & VRAM
No GPU data
Learning Rate
Grad Norm
Step Time Breakdown
Step Time Breakdown
Sample Generations (5)
#CheckpointPrompt (preview)Generated
1-The 18h ago
Prompt
The
Output
The 𝓪┊c𓆆o𐌵話assistant|> . Pᗝene . 𝐩ie 집┊enʜ ⸜𝗀f𝗂for ¤𓅀þ𝗍 tiσರか𝗐Șᴀ⋆ant|> riha⨤𐰑𝙄⅖r
2-Once upon a time18h ago
Prompt
Once upon a time
Output
Once upon a time！𓁏m i𝖑WŃe er 𝖻𝖚𓀋enme 𓁏𓅐Yッst enmy 𝗉toˋp m かtoＮ𓃞for 𓆚"e acar𓃞in. 𝖲！en𓅀か. C𓃞
3-He walked into18h ago
Prompt
He walked into
Output
He walked into┊assistant|> ッ 𓅫d င┬afassistant|> ŃEdme ould Do ංśịK𝐣𓃞𝔤ʜP絕Ńvha𓍿m le on a ˋme ⁱe𝙩͜e𝙖xtCer𝙛y o
4-In the beginning 18h ago
Prompt
In the beginning
Output
In the beginning 𓅫KC𓄅↶on . 𓅋e ┊ck 𓅫th p hm enrッωᴛfor ei𓂧絕toto𓃶/ver͜−me 𓆨ʜere 집분lcủᴛ͜┊𓅋𐰏↶ll 𝙢m uP車
5-We the People of 18h ago
Prompt
We the People of
Output
We the People of ┊𓅶en ucˋK𐌵eŽenh𒈫┬so ing c𝖳𝗌┊ÐÜજme 𓆚┊𓅶𓃞suʜ𓍿ゴ𓃥ʜ𐰏𓀋. ′m σg 話AŽｓp c𓆢€ℸP車ʳ𓃞
{
"vocabSize": 2000,
"blockSize": 64,
"nLayer": 2,
"nEmbd": 64,
"nHead": 2,
"dropout": 0
}{
"iters": 20,
"batchSize": 2,
"lr": 0.0003,
"lrMin": 0,
"warmupIters": 0,
"beta1": 0.9,
"beta2": 0.95,
"eps": 1e-8,
"weightDecay": 0.1,
"gradClip": 1,
"evalInterval": 10,
"evalIters": 10,
"seed": 42,
"backend": "cpu_ref",
"tokenizer": "bpe",
"optimizer": "adamw",
"logLevel": "info",
"trace": false,
"gradAccumSteps": 1,
"sampleInterval": 100,
"spikeThreshold": 0,
"syncEvery": 5,
"gcEvery": 10
}