passes_n_clean_d60m
d60m (68.9M parameters) on plus_all v2 subset + 39M fresh clean tokens, trained on RTX 3060 12GB (Home GPU). Started 10 Oct 2026, ended 10 Oct 2026.
Hypothesis
Reference for the OCR decision: 39M tokens of fresh clean text added to the same piece (about 20% of what the model reads) beat a fourth pass over it at the same compute.
What we learned
Yes: ex-Gita 0.6544, 1.03% better than a fourth pass (clean_v1 -0.90%), past the 0.52% noise bar, every held-out set better and a smaller train/val gap. Fresh text is worth more than another pass; the bar the OCR text must meet.
Scores
Bits per byte, lower is better; change against the parent run, Passes 4.
| Test set | Passes N-clean | Passes 4 (parent) | Change |
|---|---|---|---|
| DCS gold (classical) | 0.6655 | 0.6698 | −0.6% |
| Bhagavad-gītā (memorisation) | 0.4425 | 0.4451 | −0.6% |
| Out of domain | 0.6571 | 0.6655 | −1.3% |
| Prose | 0.6249 | 0.628 | −0.5% |
| Vedic (Ṛgveda) | 0.8862 | 0.8921 | −0.7% |
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 + 39M fresh clean tokens
- Words
- —
- Training tokens
- 194,578,466
- Passes
- 3.19 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