tokv3_v3acc_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
Adding Vedic accent marks as tokenizer symbols helps.
What we learned
No: ex-Gita 0.7187 (+0.14%), and Vedic 3.5% worse.
Scores
Bits per byte, lower is better; change against the parent run, TOK-v3 v0.1.
ex-Gītā (headline)
0.7187
bits per byte
+0.1%ex-Gītā, clean_v1
—
bits per byte
Pooled, all five sets
0.7113
bits per byte
+0.2%Validation split
0.6983
bits per byte
+0.1%| Test set | TOK-v3 v3acc | TOK-v3 v0.1 (parent) | Change |
|---|---|---|---|
| DCS gold (classical) | 0.708 | 0.7065 | +0.2% |
| Bhagavad-gītā (memorisation) | 0.5171 | 0.5129 | +0.8% |
| Out of domain | 0.7227 | 0.7225 | +0.0% |
| Prose | 0.6892 | 0.6887 | +0.1% |
| Vedic (Ṛgveda) | 1.0027 | 0.9685 | +3.5% |
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 v3acc)
Data
- Slice
- plus_clean subset
- Words
- —
- Training tokens
- 343,818,672
- Passes
- 1 passes
- Tokens seen
- 343,818,240
Compute
- GPU
- RTX 3060 12GB
- Where
- Home GPU
- Steps
- 6,995 / 6,995
- GPU hours
- 0.86
- Cost
- —
- Spot restarts
- —
Lineage
homed20m