e15_slp1_char_ctx1024
d20m (19.6M parameters) on E8 screening slice, trained on RTX 3060 12GB (Home GPU). Started 12 Sep 2026, ended 12 Sep 2026.
Hypothesis
A character-level model beats unigram when both see the same amount of text as context.
What we learned
At the standard window it loses by 0.8%; at its own 1,024-character window it is 2% better (0.7320) with the best verse completion of any 20M model, at 2.5x the tokens.
Scores
Bits per byte, lower is better; change against the parent run, E8-slp1-char.
ex-Gītā (headline)
0.7609
bits per byte
+1.4%ex-Gītā, clean_v1
—
bits per byte
Pooled, all five sets
0.7528
bits per byte
+1.5%Validation split
0.6695
bits per byte
−0.5%| Test set | E15-char-1024 | E8-slp1-char (parent) | Change |
|---|---|---|---|
| DCS gold (classical) | 0.7156 | 0.6979 | +2.5% |
| Bhagavad-gītā (memorisation) | 0.5402 | 0.5192 | +4.0% |
| Out of domain | 0.7655 | 0.7548 | +1.4% |
| Prose | 0.7379 | 0.724 | +1.9% |
| Vedic (Ṛgveda) | 1.101 | 1.1766 | −6.4% |
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
- 19,620,352
- Outside embeddings
- 18,881,024
- Layers · heads · width
- 6 · 8 · 512
- Tokenizer
- SLP1 characters
Data
- Slice
- E8 screening slice
- Words
- —
- Training tokens
- 1,660,754,802
- Passes
- 0.59 passes
- Tokens seen
- 981,811,200
Compute
- GPU
- RTX 3060 12GB
- Where
- Home GPU
- Steps
- 19,975 / 19,975
- GPU hours
- 2.12
- Cost
- —
- Spot restarts
- —
Lineage
homed20m