e8_slp1_uni8k
d20m (23.2M parameters) on E8 screening slice, trained on RTX 3060 12GB (Home GPU). Started 11 Sep 2026, ended 11 Sep 2026.
Hypothesis
SLP1 unigram 8k completes the script x algorithm grid.
What we learned
Pooled 0.7450. Became the working default and, after E9 to E14, the frozen tokenizer (SLP1 unigram 8k).
Scores
Bits per byte, lower is better.
ex-Gītā (headline)
0.7537
bits per byte
ex-Gītā, clean_v1
—
bits per byte
Pooled, all five sets
0.745
bits per byte
Validation split
0.6817
bits per byte
| Test set | E8-slp1-uni |
|---|---|
| DCS gold (classical) | 0.7161 |
| Bhagavad-gītā (memorisation) | 0.516 |
| Out of domain | 0.7544 |
| Prose | 0.726 |
| Vedic (Ṛgveda) | 1.2654 |
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 unigram 8k
Data
- Slice
- E8 screening slice
- Words
- —
- Training tokens
- 665,395,149
- Passes
- 0.59 passes
- Tokens seen
- 390,660,096
Compute
- GPU
- RTX 3060 12GB
- Where
- Home GPU
- Steps
- 7,948 / 7,948
- GPU hours
- 0.88
- Cost
- —
- Spot restarts
- —
Lineage
homed20m