e9_slp1_uni8k
d20m (22.7M parameters) on E8 screening slice, trained on RTX 3060 12GB (Home GPU). Started 12 Sep 2026, ended 12 Sep 2026.
Hypothesis
Vocabulary 8k at a fixed total parameter budget (11 layers x 384).
What we learned
Pooled 0.7466. The wider E8 arm with the same vocabulary scored 0.7450, so model shape mattered more than vocabulary here.
Scores
Bits per byte, lower is better; change against the parent run, E8-slp1-uni.
ex-Gītā (headline)
0.7552
bits per byte
+0.2%ex-Gītā, clean_v1
—
bits per byte
Pooled, all five sets
0.7466
bits per byte
+0.2%Validation split
0.6869
bits per byte
+0.8%| Test set | E9-8k | E8-slp1-uni (parent) | Change |
|---|---|---|---|
| DCS gold (classical) | 0.7191 | 0.7161 | +0.4% |
| Bhagavad-gītā (memorisation) | 0.5202 | 0.516 | +0.8% |
| Out of domain | 0.7551 | 0.7544 | +0.1% |
| Prose | 0.7295 | 0.726 | +0.5% |
| Vedic (Ṛgveda) | 1.2677 | 1.2654 | +0.2% |
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
- 22,741,632
- Outside embeddings
- 19,473,024
- Layers · heads · width
- 11 · 6 · 384
- Tokenizer
- SLP1 unigram 8k
Data
- Slice
- E8 screening slice
- Words
- —
- Training tokens
- 656,098,446
- Passes
- 0.59 passes
- Tokens seen
- 387,907,584
Compute
- GPU
- RTX 3060 12GB
- Where
- Home GPU
- Steps
- 7,892 / 7,892
- GPU hours
- 0.95
- Cost
- —
- Spot restarts
- —
Lineage
homed20m