e8_deva_bpe8k_syms
d20m (23.2M parameters) on E8 screening slice, trained on RTX 3060 12GB (Home GPU). Started 11 Sep 2026, ended 11 Sep 2026.
Hypothesis
Do dedicated verse-mark (daṇḍa) symbols in a Devanagari BPE tokenizer help a 20M model?
What we learned
No: pooled 0.7629 vs 0.7624 without the symbols, a tie. BPE trails unigram by about 2.8% at the same vocabulary.
Scores
Bits per byte, lower is better.
ex-Gītā (headline)
0.7719
bits per byte
ex-Gītā, clean_v1
—
bits per byte
Pooled, all five sets
0.7629
bits per byte
Validation split
0.6953
bits per byte
| Test set | E8-deva-bpe-sym |
|---|---|
| DCS gold (classical) | 0.7314 |
| Bhagavad-gītā (memorisation) | 0.5269 |
| Out of domain | 0.7757 |
| Prose | 0.7394 |
| Vedic (Ṛgveda) | 1.2474 |
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
- 23,239,168
- Outside embeddings
- 18,881,024
- Layers · heads · width
- 6 · 8 · 512
- Tokenizer
- Devanagari BPE 8k
Data
- Slice
- E8 screening slice
- Words
- —
- Training tokens
- 596,141,546
- Passes
- 0.59 passes
- Tokens seen
- 350,011,392
Compute
- GPU
- RTX 3060 12GB
- Where
- Home GPU
- Steps
- 7,121 / 7,121
- GPU hours
- 0.78
- Cost
- —
- Spot restarts
- —
homed20m