passes_p3_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
Reference for the passes question: a 60M model given F10's data-to-size ratio (2.25 tokens per parameter per pass over a fixed 155M-token piece of F10's own data) for three passes, as F10 had.
What we learned
Done: ex-Gita 0.6670 (clean_v1 0.6891), Gita 0.4560. Its train/val gap grows 0.16, 0.22, 0.29 nats at the pass ends, close to F10's 0.16, 0.20, 0.31, so the proxy repeats F10's memorisation pattern. The reference for Passes 4 and 5.
Scores
Bits per byte, lower is better; change against the parent run, F10.
| Test set | Passes 3 | F10 (parent) | Change |
|---|---|---|---|
| DCS gold (classical) | 0.6754 | 0.5513 | +22.5% |
| Bhagavad-gītā (memorisation) | 0.456 | 0.0758 | +501.6% |
| Out of domain | 0.671 | 0.5286 | +26.9% |
| Prose | 0.6353 | 0.5186 | +22.5% |
| Vedic (Ṛgveda) | 0.8913 | 0.6671 | +33.6% |
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
- 3 passes
- Tokens seen
- 465,813,504
Compute
- GPU
- RTX 3060 12GB
- Where
- Home GPU
- Steps
- 9,477 / 9,477
- GPU hours
- 2.88
- Cost
- —
- Spot restarts
- —
Lineage
homed60m