f3_slp1_uni8k_d125m_mix12
d125m (91.5M parameters) on rebuild 5 slice, 12.5% OCR (505.8M words, 0.81 passes), trained on RTX 3060 12GB (Home GPU). Started 19 Sep 2026, ended 20 Sep 2026.
Hypothesis
A small dose of OCR text (12.5% of tokens) helps where 45% hurt.
What we learned
No: ex-Gita 0.6336. With F2 and F2-noocr, 0%, 12.5% and 45% OCR fall on a straight line: OCR text hurts in proportion to its share.
Scores
Bits per byte, lower is better; change against the parent run, F2-noocr.
ex-Gītā (headline)
0.6336
bits per byte
+0.5%ex-Gītā, clean_v1
—
bits per byte
Pooled, all five sets
0.6235
bits per byte
+0.5%Validation split
0.636
bits per byte
+5.1%| Test set | F3-mix12 | F2-noocr (parent) | Change |
|---|---|---|---|
| DCS gold (classical) | 0.637 | 0.6319 | +0.8% |
| Bhagavad-gītā (memorisation) | 0.3597 | 0.3545 | +1.5% |
| Out of domain | 0.6298 | 0.626 | +0.6% |
| Prose | 0.6087 | 0.609 | −0.0% |
| Vedic (Ṛgveda) | 1.0848 | 1.0836 | +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, 12.5% OCR
- Words
- 505,800,000
- Training tokens
- 1,643,264,425
- Passes
- 0.81 passes
- Tokens seen
- 1,332,314,112
Compute
- GPU
- RTX 3060 12GB
- Where
- Home GPU
- Steps
- 27,106 / 27,106
- GPU hours
- 10.89
- Cost
- —
- Spot restarts
- —
Lineage
homed125m