passes_p4_d60m
d60m (68.9M parameters) on plus_all v2 subset (155M tokens), trained on RTX 3060 12GB (Home GPU). Started 7 Oct 2026, ended 7 Oct 2026.
Hypothesis
A fourth pass over the same text still improves held-out scores at F10's data-to-size ratio, rather than mostly adding memorisation.
What we learned
Better: ex-Gita 0.6612, 0.87% below three passes (clean_v1 0.6831, -0.87%), past the noise bar, for 1.33x the compute. Gita 2.4% lower (more memorised), Vedic flat. Recommended for the next 700M run.
Scores
Bits per byte, lower is better; change against the parent run, Passes 3.
| Test set | Passes 4 | Passes 3 (parent) | Change |
|---|---|---|---|
| DCS gold (classical) | 0.6698 | 0.6754 | −0.8% |
| Bhagavad-gītā (memorisation) | 0.4451 | 0.456 | −2.4% |
| Out of domain | 0.6655 | 0.671 | −0.8% |
| Prose | 0.628 | 0.6353 | −1.1% |
| Vedic (Ṛgveda) | 0.8921 | 0.8913 | +0.1% |
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
- 68,924,160
- Outside embeddings
- 56,636,160
- Layers · heads · width
- 8 · 12 · 768
- Tokenizer
- SLP1 unigram 8k
Data
- Slice
- plus_all v2 subset (155M tokens)
- Words
- —
- Training tokens
- 155,261,295
- Passes
- 4 passes
- Tokens seen
- 621,084,672
Compute
- GPU
- RTX 3060 12GB
- Where
- Home GPU
- Steps
- 12,636 / 12,636
- GPU hours
- 3.85
- Cost
- —
- Spot restarts
- —
Lineage
homed60m