f0_slp1_uni8k_d60m
d60m (63.2M parameters) on rebuild 3 slice (414M words, 1 passes), trained on RTX 3060 12GB (Home GPU). Started 13 Sep 2026, ended 13 Sep 2026.
Hypothesis
The first full-corpus run (414M words, one pass) takes a 60M model well below the screening models.
What we learned
Pooled 0.6282, 10.5% better than E14. Held-out Gita verses turned out to sit inside training texts, so the Gita column measures memorisation; masking came next.
Scores
Bits per byte, lower is better; change against the parent run, E14-8k.
ex-Gītā (headline)
0.6434
bits per byte
−9.6%ex-Gītā, clean_v1
—
bits per byte
Pooled, all five sets
0.6282
bits per byte
−10.5%Validation split
0.6261
bits per byte
−3.2%| Test set | F0 | E14-8k (parent) | Change |
|---|---|---|---|
| DCS gold (classical) | 0.6344 | 0.6737 | −5.8% |
| Bhagavad-gītā (memorisation) | 0.2284 | 0.4374 | −47.8% |
| Out of domain | 0.639 | 0.7097 | −10.0% |
| Prose | 0.6238 | 0.6868 | −9.2% |
| Vedic (Ṛgveda) | 1.1091 | 1.292 | −14.2% |
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
- 63,173,376
- Outside embeddings
- 56,636,160
- Layers · heads · width
- 8 · 12 · 768
- Tokenizer
- SLP1 unigram 8k
Data
- Slice
- rebuild 3 slice
- Words
- 414,000,000
- Training tokens
- 1,337,088,903
- Passes
- 1 passes
- Tokens seen
- 1,337,081,856
Compute
- GPU
- RTX 3060 12GB
- Where
- Home GPU
- Steps
- 27,203 / 27,203
- GPU hours
- 7.49
- Cost
- —
- Spot restarts
- —
Lineage
homed60m