ocrvlm_v12_clean_q35
on OCR page labels, clean, trained on A100 40GB (Google Cloud, spot). Started 8 Oct 2026, ended 8 Oct 2026.
Hypothesis
Retrained without any labels from Google Vision, Claude, Codex or Sarvam, the same 2B OCR model still beats Google Vision's 1.69% letter error rate.
What we learned
Yes: median letter error 1.06% on 54 held-out pages against Google Vision's 1.69%, better than Vision on 39 of 54 pages. OCR-VLM v1, trained with those services' labels, scored 0.96%: the clean labels cost about a tenth of a point. This is the released OCR model.
Scores
Lower is better on the headline; change against the parent run, OCR-VLM v1.
Letter error rate: the share of letters that come out wrong, as a fraction (0.0096 means 0.96% of letters): for OCR, letters read off a page; for a voice, letters a speech recogniser hears in its audio. Lower is better.
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
- —
- Parameters
- —
- Outside embeddings
- —
- Layers · heads · width
- — · — · —
- Tokenizer
- Qwen3.5 tokenizer
Data
- Slice
- OCR page labels, clean
- Words
- —
- Training tokens
- —
- Passes
- —
- Tokens seen
- 80,725,777
Compute
- GPU
- A100 40GB
- Where
- Google Cloud (spot)
- Steps
- 2,000 / 2,000
- GPU hours
- 7.59
- Cost
- $9.86
- Spot restarts
- 1
Lineage
gcpa100spot