stale
fine_corpus_l4_20260308_061132
20.51M parameter fine_corpus model — bpe-8k tokenizer, 8L/384D/6H
Overview
20.51M
Parameters
7.8845
Final Loss
7.9634
Best Val Loss
2655.8
Perplexity
52,451,328
Tokens Processed
2.6
Tokens/Param
4,716 tok/s
Avg Throughput
1h 14m
Training Time
Training Progress25,611 / 78,000 steps (32.8%)
Loss reduced by 13.0% from initial 9.0639
Dataset & Training
Domainfine_corpus
Tokenizerbpe-8k
Total Iterations78,000
Batch Size4
Context Length512 tokens
Tokens per Batch2,048
Dataset Passes~66
Effective Tokens52,451,328
Training Pipeline
Warmupsteps 1–780
Learning rate warmup — model weights adjusting to data distribution
Loss: 9.064 → 7.943Linear LR warmup, gradient clipping
Training Metrics
Loss Curve
?
?
?
?
Smoothed Loss
Perplexity
Learning Rate
Gradient Norm
Throughput (tok/s)
Timing Breakdown
No Telemetry
Model Architecture
Model Configuration
ArchitectureGPT (decoder-only transformer)
Parameters20.51M
Layers8
Embedding Dim384
Attention Heads6
Head Dim64
FFN Dim1536
FFN Activationgelu
Vocab Size8,000
Context Length512 tokens
Dropout0
Training Configuration
Optimizeradamw
Learning Rate0.0006
LR Min0.00006
LR ScheduleCosine decay
Warmup Steps500
Batch Size4
Grad Accum Steps1
Effective Batch4
Grad Clip1
Weight Decay0.1
Backendhelios
Tokenizerbpe-8k
Seed42
Layer Structure
Token Embed
8,000×384
Pos Embed
512×384
Block 0
Attn+FFN
Block 1
Attn+FFN
Block 2
Attn+FFN
Block 3
Attn+FFN
Block 4
Attn+FFN
Block 5
Attn+FFN
LayerNorm
384
LM Head
384×8,000
Generated Samples
Step 0 — Mar 8, 2026 9:26 AM
Prompt: The
The 进西直ؕEinstein's primorCornu守ُCornu精ٌٍُ♀
Checkpoints
No checkpoints saved yet.
Chat with Model
Send a message to chat with this model
Generated Invalid Date Invalid Date — Alpha Training SystemConfig hash: