kg_e11_slp1_bpe8k_dropout
d20m (23.2M parameters) on E8 screening slice, trained on T4 16GB (Kaggle). Started 12 Sep 2026, ended 12 Sep 2026.
Hypothesis
Randomly varying the BPE segmentation during training (BPE-dropout) regularises the model.
What we learned
Hurt by 5.8% (pooled 0.8059 vs 0.7619): the noise is paid for in training and never returned at evaluation, where segmentation is fixed.
Scores
Bits per byte, lower is better; change against the parent run, E8-slp1-bpe.
ex-Gītā (headline)
0.814
bits per byte
+5.6%ex-Gītā, clean_v1
—
bits per byte
Pooled, all five sets
0.8059
bits per byte
+5.8%Validation split
0.7299
bits per byte
+5.4%| Test set | E11-bpe-dropout | E8-slp1-bpe (parent) | Change |
|---|---|---|---|
| DCS gold (classical) | 0.776 | 0.7355 | +5.5% |
| Bhagavad-gītā (memorisation) | 0.5936 | 0.5313 | +11.7% |
| Out of domain | 0.8164 | 0.7713 | +5.8% |
| Prose | 0.7841 | 0.7417 | +5.7% |
| Vedic (Ṛgveda) | 1.2958 | 1.2938 | +0.2% |
Curves
Training loss
Cross-entropy per token, by step. The first few percent of the run, far higher, run off the top; hover or the table has every value.
Held-out bits per byte
The validation split, evaluated during training, by step. Lower is better.
Throughput
Tokens per second, by step.
Model
- Preset
- d20m
- Parameters
- 23,239,168
- Outside embeddings
- 18,881,024
- Layers · heads · width
- 6 · 8 · 512
- Tokenizer
- SLP1 BPE 8k
Data
- Slice
- E8 screening slice
- Words
- —
- Training tokens
- 690,634,473
- Passes
- 0.59 passes
- Tokens seen
- 408,256,512
Compute
- GPU
- T4 16GB
- Where
- Kaggle
- Steps
- 8,306 / 8,306
- GPU hours
- 1.65
- Cost
- Free
- Spot restarts
- —
Lineage
kagglet4d20m