e9_slp1_uni32k
d20m (23.1M parameters) on E8 screening slice, trained on RTX 3060 12GB (Home GPU). Started 12 Sep 2026, ended 12 Sep 2026.
Hypothesis
Vocabulary 32k at a fixed total parameter budget.
What we learned
Pooled 0.7576, the worst of E9: a 32k table eats 54% of the budget and leaves only 6 layers.
Scores
Bits per byte, lower is better; change against the parent run, E8-slp1-uni.
ex-Gītā (headline)
0.7671
bits per byte
+1.8%ex-Gītā, clean_v1
—
bits per byte
Pooled, all five sets
0.7576
bits per byte
+1.7%Validation split
0.6889
bits per byte
+1.1%| Test set | E9-32k | E8-slp1-uni (parent) | Change |
|---|---|---|---|
| DCS gold (classical) | 0.7187 | 0.7161 | +0.4% |
| Bhagavad-gītā (memorisation) | 0.5072 | 0.516 | −1.7% |
| Out of domain | 0.7703 | 0.7544 | +2.1% |
| Prose | 0.7342 | 0.726 | +1.1% |
| Vedic (Ṛgveda) | 1.3075 | 1.2654 | +3.3% |
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
- 23,106,432
- Outside embeddings
- 10,621,824
- Layers · heads · width
- 6 · 6 · 384
- Tokenizer
- SLP1 unigram 32k
Data
- Slice
- E8 screening slice
- Words
- —
- Training tokens
- 544,728,329
- Passes
- 0.59 passes
- Tokens seen
- 322,043,904
Compute
- GPU
- RTX 3060 12GB
- Where
- Home GPU
- Steps
- 6,552 / 6,552
- GPU hours
- 0.76
- Cost
- —
- Spot restarts
- —
Lineage
homed20m