visiontest_c_s1338_d20m
d20m (27.1M parameters) on plus_clean subset, trained on RTX 3060 12GB (Home GPU). Started 3 Oct 2026, ended 3 Oct 2026.
Hypothesis
The same 12.5% taken from the old scanned-book OCR of the same nine books.
What we learned
1.1% worse than clean text; Vision and old OCR tie.
Scores
Bits per byte, lower is better; change against the parent run, VISION-OCR A.
ex-Gītā (headline)
0.8002
bits per byte
+1.1%ex-Gītā, clean_v1
—
bits per byte
Pooled, all five sets
0.7945
bits per byte
+1.1%Validation split
0.8127
bits per byte
+3.8%| Test set | VISION-OCR C s2 | VISION-OCR A (parent) | Change |
|---|---|---|---|
| DCS gold (classical) | 0.7841 | 0.7758 | +1.1% |
| Bhagavad-gītā (memorisation) | 0.645 | 0.6338 | +1.8% |
| Out of domain | 0.8068 | 0.7976 | +1.2% |
| Prose | 0.7638 | 0.756 | +1.0% |
| Vedic (Ṛgveda) | 1.1011 | 1.0937 | +0.7% |
Curves
No training curves were published for this run.Its scores above are what it published.
Model
- Preset
- d20m
- Parameters
- 27,073,024
- Outside embeddings
- 18,881,024
- Layers · heads · width
- 6 · 8 · 512
- Tokenizer
- SLP1 unigram 8k
Data
- Slice
- plus_clean subset
- Words
- —
- Training tokens
- 90,848,863
- Passes
- 1 passes
- Tokens seen
- 90,783,744
Compute
- GPU
- RTX 3060 12GB
- Where
- Home GPU
- Steps
- 1,847 / 1,847
- GPU hours
- 0.23
- Cost
- —
- Spot restarts
- —
Lineage
homed20m