f2_slp1_uni8k_d125m_noocr
d125m (91.5M parameters) on rebuild 5 slice without OCR (446.9M words, 0.93 passes), trained on RTX 3060 12GB (Home GPU). Started 19 Sep 2026, ended 19 Sep 2026.
Hypothesis
Dropping the scanned-book OCR text recovers what F2 lost.
What we learned
Yes: ex-Gita 0.6306, 1.0% better than F1b and 1.5% better than F2, the first clear data win. The new clean text helps; the OCR text cancelled it.
Scores
Bits per byte, lower is better; change against the parent run, F2.
ex-Gītā (headline)
0.6306
bits per byte
−1.5%ex-Gītā, clean_v1
—
bits per byte
Pooled, all five sets
0.6205
bits per byte
−1.4%Validation split
0.6053
bits per byte
−14.1%| Test set | F2-noocr | F2 (parent) | Change |
|---|---|---|---|
| DCS gold (classical) | 0.6319 | 0.6515 | −3.0% |
| Bhagavad-gītā (memorisation) | 0.3545 | 0.3571 | −0.7% |
| Out of domain | 0.626 | 0.636 | −1.6% |
| Prose | 0.609 | 0.6139 | −0.8% |
| Vedic (Ṛgveda) | 1.0836 | 1.0752 | +0.8% |
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
- 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 without OCR
- Words
- 446,900,000
- Training tokens
- 1,438,157,255
- Passes
- 0.93 passes
- Tokens seen
- 1,332,314,112
Compute
- GPU
- RTX 3060 12GB
- Where
- Home GPU
- Steps
- 27,106 / 27,106
- GPU hours
- 10.97
- Cost
- —
- Spot restarts
- —
Lineage
homed125m