passes_p2_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
Two passes: the low end of the passes curve.
What we learned
Worse: ex-Gita 0.6773, 1.55% behind three passes (clean_v1 0.6990, +1.44%), Vedic 3.7% worse. The third pass still pays.
Scores
Bits per byte, lower is better; change against the parent run, Passes 3.
ex-Gītā (headline)
0.6773
bits per byte
+1.5%ex-Gītā, clean_v1
0.699
bits per byte
+1.4%Pooled, all five sets
0.67
bits per byte
+1.6%Validation split
0.6699
bits per byte
+1.7%| Test set | Passes 2 | Passes 3 (parent) | Change |
|---|---|---|---|
| DCS gold (classical) | 0.6873 | 0.6754 | +1.8% |
| Bhagavad-gītā (memorisation) | 0.4784 | 0.456 | +4.9% |
| Out of domain | 0.6813 | 0.671 | +1.5% |
| Prose | 0.6434 | 0.6353 | +1.3% |
| Vedic (Ṛgveda) | 0.9243 | 0.8913 | +3.7% |
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 (155M tokens)
- Words
- —
- Training tokens
- 155,261,295
- Passes
- 2 passes
- Tokens seen
- 310,542,336
Compute
- GPU
- RTX 3060 12GB
- Where
- Home GPU
- Steps
- 6,318 / 6,318
- GPU hours
- 1.92
- Cost
- —
- Spot restarts
- —
Lineage
homed60m