passes_p2_muonio_d60m
d60m (68.9M parameters) on plus_all v2 subset (155M tokens), trained on RTX 3060 12GB (Home GPU). Started 11 Oct 2026, ended 11 Oct 2026.
Hypothesis
Updating the word-embedding and output tables with a per-token normalised rule (MuonIO, a 2026 optimizer paper) instead of AdamW gives a better four-pass model.
What we learned
Diverged with the paper's 60M settings: the rule grows the tables' rows to its own scale (about 29 here against 0.4 normally) and the loss climbed from step 150, so it was stopped at step 500. A rescaled rerun follows.
Scores
Bits per byte, lower is better; change against the parent run, Passes 4.
| Test set | Passes MuonIO | Passes 4 (parent) | Change |
|---|---|---|---|
| DCS gold (classical) | — | 0.6698 | |
| Bhagavad-gītā (memorisation) | — | 0.4451 | |
| Out of domain | — | 0.6655 | |
| Prose | — | 0.628 | |
| Vedic (Ṛgveda) | — | 0.8921 |
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
- 0.16 passes
- Tokens seen
- 24,576,000
Compute
- GPU
- RTX 3060 12GB
- Where
- Home GPU
- Steps
- 500 / 12,636
- GPU hours
- 0.15
- Cost
- —
- Spot restarts
- —
Lineage
homed60m