e8_slp1_char
d20m (19.4M parameters) on E8 screening slice, trained on RTX 3060 12GB (Home GPU). Started 11 Sep 2026, ended 11 Sep 2026.
Hypothesis
Character-level models are the weakest option at 20M parameters.
What we learned
Wrong: pooled 0.7419 ties unigram and wins Vedic by 5.6%, at 2.4x the compute and a quarter of the context. Reopened as E15.
Scores
Bits per byte, lower is better.
ex-Gītā (headline)
0.7503
bits per byte
ex-Gītā, clean_v1
—
bits per byte
Pooled, all five sets
0.7419
bits per byte
Validation split
0.6732
bits per byte
| Test set | E8-slp1-char |
|---|---|
| DCS gold (classical) | 0.6979 |
| Bhagavad-gītā (memorisation) | 0.5192 |
| Out of domain | 0.7548 |
| Prose | 0.724 |
| Vedic (Ṛgveda) | 1.1766 |
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
- 19,358,208
- Outside embeddings
- 18,881,024
- Layers · heads · width
- 6 · 8 · 512
- Tokenizer
- SLP1 characters
Data
- Slice
- E8 screening slice
- Words
- —
- Training tokens
- 1,673,945,605
- Passes
- 0.59 passes
- Tokens seen
- 982,794,240
Compute
- GPU
- RTX 3060 12GB
- Where
- Home GPU
- Steps
- 19,995 / 19,995
- GPU hours
- 1.85
- Cost
- —
- Spot restarts
- —
Lineage
homed20m