f2_slp1_uni8k_d125m
d125m (91.5M parameters) on rebuild 5 slice (798.3M words, 0.5 passes), trained on RTX 3060 12GB (Home GPU). Started 18 Sep 2026, ended 19 Sep 2026.
Hypothesis
A slice twice as large (798M words, about 45% of tokens from scanned-book OCR) improves held-out quality at equal compute.
What we learned
No: ex-Gita 0.6399 vs 0.6372, inside the noise, and the cleanest set got 2.9% worse.
Scores
Bits per byte, lower is better; change against the parent run, F1b.
ex-Gītā (headline)
0.6399
bits per byte
+0.4%ex-Gītā, clean_v1
—
bits per byte
Pooled, all five sets
0.6296
bits per byte
+0.4%Validation split
0.7045
bits per byte
+15.0%| Test set | F2 | F1b (parent) | Change |
|---|---|---|---|
| DCS gold (classical) | 0.6515 | 0.633 | +2.9% |
| Bhagavad-gītā (memorisation) | 0.3571 | 0.3597 | −0.7% |
| Out of domain | 0.636 | 0.6351 | +0.1% |
| Prose | 0.6139 | 0.6114 | +0.4% |
| Vedic (Ṛgveda) | 1.0752 | 1.0743 | +0.1% |
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
- d125m
- Parameters
- 91,491,072
- Outside embeddings
- 84,953,856
- Layers · heads · width
- 12 · 12 · 768
- Tokenizer
- SLP1 unigram 8k
Data
- Slice
- rebuild 5 slice
- Words
- 798,300,000
- Training tokens
- 2,663,420,147
- Passes
- 0.5 passes
- Tokens seen
- 1,332,314,112
Compute
- GPU
- RTX 3060 12GB
- Where
- Home GPU
- Steps
- 27,106 / 27,106
- GPU hours
- 10.95
- Cost
- —
- Spot restarts
- —
Lineage
homed125m