f10ab_a_d60m
d60m (68.9M parameters) on plus_all band (96M words, 1 passes), trained on RTX 3060 12GB (Home GPU). Started 3 Oct 2026, ended 3 Oct 2026.
Hypothesis
Control for the quality filters: a 96M-word band of the F9 slice with only held-out removal.
What we learned
ex-Gita 0.6831, the reference for arm b.
Scores
Bits per byte, lower is better; change against the parent run, F9.
ex-Gītā (headline)
0.6831
bits per byte
+23.1%ex-Gītā, clean_v1
—
bits per byte
Pooled, all five sets
0.6754
bits per byte
+25.2%Validation split
0.6628
bits per byte
+31.7%| Test set | F10-AB a | F9 (parent) | Change |
|---|---|---|---|
| DCS gold (classical) | 0.6893 | 0.5687 | +21.2% |
| Bhagavad-gītā (memorisation) | 0.4724 | 0.1367 | +245.6% |
| Out of domain | 0.6831 | 0.5542 | +23.3% |
| Prose | 0.6564 | 0.5406 | +21.4% |
| Vedic (Ṛgveda) | 1.0048 | 0.6892 | +45.8% |
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 band
- Words
- 96,000,000
- Training tokens
- 312,487,176
- Passes
- 1 passes
- Tokens seen
- 312,508,416
Compute
- GPU
- RTX 3060 12GB
- Where
- Home GPU
- Steps
- 6,358 / 6,358
- GPU hours
- 1.91
- Cost
- —
- Spot restarts
- —
Lineage
homed60m