e9_slp1_uni16k
d20m (24M parameters) on E8 screening slice, trained on RTX 3060 12GB (Home GPU). Started 12 Sep 2026, ended 12 Sep 2026.
Hypothesis
Vocabulary 16k at a fixed total parameter budget (10 layers, 26% of parameters in embeddings).
What we learned
Pooled 0.7465, a tie with 8k: vocabulary size is flat between 8k and 16k at this scale.
Scores
Bits per byte, lower is better; change against the parent run, E8-slp1-uni.
ex-Gītā (headline)
0.7556
bits per byte
+0.3%ex-Gītā, clean_v1
—
bits per byte
Pooled, all five sets
0.7465
bits per byte
+0.2%Validation split
0.6805
bits per byte
−0.2%| Test set | E9-16k | E8-slp1-uni (parent) | Change |
|---|---|---|---|
| DCS gold (classical) | 0.7139 | 0.7161 | −0.3% |
| Bhagavad-gītā (memorisation) | 0.5075 | 0.516 | −1.6% |
| Out of domain | 0.7584 | 0.7544 | +0.5% |
| Prose | 0.7246 | 0.726 | −0.2% |
| Vedic (Ṛgveda) | 1.2575 | 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
- 24,043,392
- Outside embeddings
- 17,702,784
- Layers · heads · width
- 10 · 6 · 384
- Tokenizer
- SLP1 unigram 16k
Data
- Slice
- E8 screening slice
- Words
- —
- Training tokens
- 594,539,065
- Passes
- 0.59 passes
- Tokens seen
- 351,485,952
Compute
- GPU
- RTX 3060 12GB
- Where
- Home GPU
- Steps
- 7,151 / 7,151
- GPU hours
- 0.91
- Cost
- —
- Spot restarts
- —
Lineage
homed20m