f10ab_b_d60m
d60m (68.9M parameters) on plus_all band, filtered (90M words, 1.07 passes), trained on RTX 3060 12GB (Home GPU). Started 3 Oct 2026, ended 4 Oct 2026.
Hypothesis
Quality filters (transliteration junk, other languages, garbled text) on the same band help.
What we learned
Tie: ex-Gita 0.6814 vs 0.6831 (-0.26%, inside the noise). Dropping 6% of the words costs nothing, so F10 trains on the filtered slice.
Scores
Bits per byte, lower is better; change against the parent run, F10-AB a.
ex-Gītā (headline)
0.6814
bits per byte
−0.2%ex-Gītā, clean_v1
—
bits per byte
Pooled, all five sets
0.6735
bits per byte
−0.3%Validation split
0.658
bits per byte
−0.7%| Test set | F10-AB b | F10-AB a (parent) | Change |
|---|---|---|---|
| DCS gold (classical) | 0.686 | 0.6893 | −0.5% |
| Bhagavad-gītā (memorisation) | 0.4656 | 0.4724 | −1.4% |
| Out of domain | 0.6819 | 0.6831 | −0.2% |
| Prose | 0.6538 | 0.6564 | −0.4% |
| Vedic (Ṛgveda) | 0.9998 | 1.0048 | −0.5% |
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, filtered
- Words
- 90,000,000
- Training tokens
- 291,081,028
- Passes
- 1.07 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