visiontest_e_s1337_d20m
d20m (27.1M parameters) on plus_clean subset, trained on RTX 3060 12GB (Home GPU). Started 10 Oct 2026, ended 10 Oct 2026.
Hypothesis
Text from our own open OCR model (sansar-ocr-2b) is as good as clean e-text: the same nine books read by it cost within 0.2% of arm D.
What we learned
No: 0.4% worse than arm D (seed mean; +0.33% and +0.42%), against a 0.2% bar. It ties Google Vision and the old OCR, and is worst on Vedic. Like them, it behaves as an OCR tier at this size. Internal test.
Scores
Bits per byte, lower is better; change against the parent run, VISION-OCR A.
ex-Gītā (headline)
0.8019
bits per byte
+1.3%ex-Gītā, clean_v1
—
bits per byte
Pooled, all five sets
0.7962
bits per byte
+1.3%Validation split
0.7694
bits per byte
−1.8%| Test set | VISION-OCR E | VISION-OCR A (parent) | Change |
|---|---|---|---|
| DCS gold (classical) | 0.7844 | 0.7758 | +1.1% |
| Bhagavad-gītā (memorisation) | 0.6467 | 0.6338 | +2.0% |
| Out of domain | 0.8093 | 0.7976 | +1.5% |
| Prose | 0.7626 | 0.756 | +0.9% |
| Vedic (Ṛgveda) | 1.1247 | 1.0937 | +2.8% |
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
- 87,975,961
- Passes
- 1.01 passes
- Tokens seen
- 88,473,600
Compute
- GPU
- RTX 3060 12GB
- Where
- Home GPU
- Steps
- 1,800 / 1,800
- GPU hours
- 0.23
- Cost
- —
- Spot restarts
- —
Lineage
homed20m