Training machine offline · last report 17 days ago

OCR training runs

Reading scanned Sanskrit books: an open vision-language model fine-tuned for page OCR, and the bulk OCR jobs that turn scans into corpus text.

Headline score: Median letter error rate · lower is better The share of letters read wrong on 54 held-out book pages, median over pages: 0.0096 means 0.96% of letters.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.

Progress

Lower is better on the headline.

Best Median letter error rate over time

Each dot is a scored run, placed at the day it ended; the line is the best score so far. Numbered lines are this track's milestones, listed below.

Milestones
  1. 9 Oct 2026 · Sansar OCR 2B public on Hugging Face (OCR-VLM v1.2 clean)
  2. 6 Oct 2026 · OCR-VLM v1 beats Google Vision on Sanskrit pages
  3. 5 Oct 2026 · OCR-VLM pilot: first OCR run

OCR runs

Published · 5 runs

All OCR runs. Column headers sort the table.
Data
OCR-VLM v1.2 clean
ocrvlm_v12_clean_q35
8 Oct 2026 — OCR page labels, clean 7.59 $9.86 Winner 0.0106
OCR-VLM v1
ocrvlm_v1_q35
5 Oct 2026 — OCR page labels 7.33 $9.53 Winner 0.0096

Cost “—” means no cloud bill (our own GPU at home) or a cost that was not recorded; “Free” is free cloud compute. GPU-h is GPU hours.