e9_slp1_uni4k
d20m (23M parameters) on E8 screening slice, trained on RTX 3060 12GB (Home GPU). Started 12 Sep 2026, ended 12 Sep 2026.
Hypothesis
At a fixed total size, a smaller 4k vocabulary buys extra depth and wins.
What we learned
No: pooled 0.7497 vs 0.7466 for 8k. The smaller table does not pay off.
Scores
Bits per byte, lower is better; change against the parent run, E8-slp1-uni.
ex-Gītā (headline)
0.7581
bits per byte
+0.6%ex-Gītā, clean_v1
—
bits per byte
Pooled, all five sets
0.7497
bits per byte
+0.6%Validation split
0.6818
bits per byte
+0.0%| Test set | E9-4k | E8-slp1-uni (parent) | Change |
|---|---|---|---|
| DCS gold (classical) | 0.7238 | 0.7161 | +1.1% |
| Bhagavad-gītā (memorisation) | 0.5289 | 0.516 | +2.5% |
| Out of domain | 0.7573 | 0.7544 | +0.4% |
| Prose | 0.7335 | 0.726 | +1.0% |
| Vedic (Ṛgveda) | 1.2728 | 1.2654 | +0.6% |
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
- 22,975,872
- Outside embeddings
- 21,243,264
- Layers · heads · width
- 12 · 6 · 384
- Tokenizer
- SLP1 unigram 4k
Data
- Slice
- E8 screening slice
- Words
- —
- Training tokens
- 727,161,977
- Passes
- 0.59 passes
- Tokens seen
- 429,883,392
Compute
- GPU
- RTX 3060 12GB
- Where
- Home GPU
- Steps
- 8,746 / 8,746
- GPU hours
- 1.07
- Cost
- —
- Spot restarts
- —
Lineage
homed20m