i9small_vocab16k_d60m
d60m (81.2M parameters) on plus_clean subset, trained on RTX 3060 12GB (Home GPU). Started 3 Oct 2026, ended 3 Oct 2026.
Hypothesis
A 16k vocabulary beats 8k at 60M with the non-embedding size fixed.
What we learned
Tie: ex-Gita 0.6843 vs 0.6825 (+0.14%). A third scale point agreeing with E9 and E14, so 8k stays.
Scores
Bits per byte, lower is better; change against the parent run, E-CTX base.
ex-Gītā (headline)
0.6843
bits per byte
+0.3%ex-Gītā, clean_v1
—
bits per byte
Pooled, all five sets
0.6757
bits per byte
+0.2%Validation split
0.6609
bits per byte
−1.2%| Test set | E-VOCAB 16k | E-CTX base (parent) | Change |
|---|---|---|---|
| DCS gold (classical) | 0.6785 | 0.6795 | −0.1% |
| Bhagavad-gītā (memorisation) | 0.4506 | 0.4622 | −2.5% |
| Out of domain | 0.6891 | 0.686 | +0.5% |
| Prose | 0.6536 | 0.6547 | −0.2% |
| Vedic (Ṛgveda) | 0.9342 | 0.9254 | +1.0% |
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
- d60m
- Parameters
- 81,212,160
- Outside embeddings
- 56,636,160
- Layers · heads · width
- 8 · 12 · 768
- Tokenizer
- SLP1 unigram 16k
Data
- Slice
- plus_clean subset
- Words
- —
- Training tokens
- 294,580,283
- Passes
- 1 passes
- Tokens seen
- 294,567,936
Compute
- GPU
- RTX 3060 12GB
- Where
- Home GPU
- Steps
- 5,993 / 5,993
- GPU hours
- 1.95
- Cost
- —
- Spot restarts
- —
Lineage
homed60m