kg_slp1_uni16k_d20m
d20m (27.3M parameters) on E8 screening slice, trained on T4 16GB (Kaggle). Started 12 Sep 2026, ended 12 Sep 2026.
Hypothesis
Unigram beats BPE at a 16k vocabulary too.
What we learned
Pooled 0.7536, 1.8% better than BPE 16k.
Scores
Bits per byte, lower is better; change against the parent run, E8-slp1-uni.
ex-Gītā (headline)
0.7626
bits per byte
+1.2%ex-Gītā, clean_v1
—
bits per byte
Pooled, all five sets
0.7536
bits per byte
+1.2%Validation split
0.6872
bits per byte
+0.8%| Test set | E10-uni16k | E8-slp1-uni (parent) | Change |
|---|---|---|---|
| DCS gold (classical) | 0.7202 | 0.7161 | +0.6% |
| Bhagavad-gītā (memorisation) | 0.5162 | 0.516 | +0.0% |
| Out of domain | 0.7656 | 0.7544 | +1.5% |
| Prose | 0.7308 | 0.726 | +0.7% |
| Vedic (Ṛgveda) | 1.2688 | 1.2654 | +0.3% |
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
- 27,335,168
- Outside embeddings
- 18,881,024
- Layers · heads · width
- 6 · 8 · 512
- Tokenizer
- SLP1 unigram 16k
Data
- Slice
- E8 screening slice
- Words
- —
- Training tokens
- 594,539,065
- Passes
- 0.59 passes
- Tokens seen
- 351,485,952
Compute
- GPU
- T4 16GB
- Where
- Kaggle
- Steps
- 7,151 / 7,151
- GPU hours
- 1.41
- Cost
- Free
- Spot restarts
- —
Lineage
kagglet4d20m