tokv3_v3data_d20m
d20m (27.1M parameters) on plus_clean subset, trained on RTX 3060 12GB (Home GPU). Started 2 Oct 2026, ended 2 Oct 2026.
Hypothesis
A tokenizer retrained on today's cleaner corpus helps.
What we learned
No: ex-Gita 0.7191 (+0.20%). It packs 4.2% more text per token, which does not show up as better bits per byte. v0.1 kept.
Scores
Bits per byte, lower is better; change against the parent run, TOK-v3 v0.1.
ex-Gītā (headline)
0.7191
bits per byte
+0.2%ex-Gītā, clean_v1
—
bits per byte
Pooled, all five sets
0.7118
bits per byte
+0.2%Validation split
0.6979
bits per byte
+0.1%| Test set | TOK-v3 v3data | TOK-v3 v0.1 (parent) | Change |
|---|---|---|---|
| DCS gold (classical) | 0.7113 | 0.7065 | +0.7% |
| Bhagavad-gītā (memorisation) | 0.5182 | 0.5129 | +1.0% |
| Out of domain | 0.7223 | 0.7225 | −0.0% |
| Prose | 0.691 | 0.6887 | +0.3% |
| Vedic (Ṛgveda) | 1.0071 | 0.9685 | +4.0% |
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
- d20m
- Parameters
- 27,073,024
- Outside embeddings
- 18,881,024
- Layers · heads · width
- 6 · 8 · 512
- Tokenizer
- SLP1 unigram 8k (candidate v3data)
Data
- Slice
- plus_clean subset
- Words
- —
- Training tokens
- 328,779,352
- Passes
- 1 passes
- Tokens seen
- 328,777,728
Compute
- GPU
- RTX 3060 12GB
- Where
- Home GPU
- Steps
- 6,689 / 6,689
- GPU hours
- 0.82
- Cost
- —
- Spot restarts
- —
Lineage
homed20m