kg_slp1_bpe16k
d20m (27.3M parameters) on E8 screening slice, trained on T4 16GB (Kaggle). Started 12 Sep 2026, ended 12 Sep 2026.
Hypothesis
BPE at a 16k vocabulary.
What we learned
Pooled 0.7672: loses to unigram 16k by 1.8% and to BPE 8k. Unigram stays.
Scores
Bits per byte, lower is better; change against the parent run, E8-slp1-bpe.
ex-Gītā (headline)
0.7762
bits per byte
+0.7%ex-Gītā, clean_v1
—
bits per byte
Pooled, all five sets
0.7672
bits per byte
+0.7%Validation split
0.6976
bits per byte
+0.7%| Test set | E10-bpe16k | E8-slp1-bpe (parent) | Change |
|---|---|---|---|
| DCS gold (classical) | 0.7342 | 0.7355 | −0.2% |
| Bhagavad-gītā (memorisation) | 0.5306 | 0.5313 | −0.1% |
| Out of domain | 0.7795 | 0.7713 | +1.1% |
| Prose | 0.7452 | 0.7417 | +0.5% |
| Vedic (Ṛgveda) | 1.2587 | 1.2938 | −2.7% |
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 BPE 16k
Data
- Slice
- E8 screening slice
- Words
- —
- Training tokens
- 535,553,251
- Passes
- 0.59 passes
- Tokens seen
- 316,588,032
Compute
- GPU
- T4 16GB
- Where
- Kaggle
- Steps
- 6,441 / 6,441
- GPU hours
- 1.28
- Cost
- Free
- Spot restarts
- —
Lineage
kagglet4d20m