Tibetan OCR Leaderboard

OCR systems ranked on the open 472-page Tibetan OCR benchmark created by BDRC (Buddhist Digital Resource Center). This is a fair generalist board: every model is scored on all pages, and any page a model skips (including the pages script-specialist models deliberately avoid) counts as a full error. Lower Character Error Rate (CER) is better. CER is normalized per page and capped at 100% — a page can never count as more than fully wrong, so runaway/repetition output can't distort the average.

📚 Benchmark dataset

How to read a bar

Each bar summarises one model's per-page CER distribution across all pages — not just a single average. The x-axis is CER from 0 % (perfect) on the left to 100 % (fully wrong) on the right. Further left is better.

Whiskers — the 10th to 90th percentile of pages.
Box — the middle half of pages (25th–75th percentile).
Median line — the typical page (50th percentile).
Mean dot — the average, pulled right by the worst pages. This is what the ranking uses.