Leaderboards

AI models and skills for Buddhist tasks, ranked on open benchmarks. Each task board compares skills across models, with every model also tested on its own (“No skill”). Question-and-answer benchmarks and other open leaderboards are listed alongside. The top shippable skill for each task and language is the one people install from buddhist-skills.

Tibetan · structure_f1

Table-of-contents extraction

Benchmark toc-extraction/bo 2.0.0 · 85 items

Top performance

TOC extraction (sa-bcad method) on Sonnet 5.5

0.962≈$1.20 per 100 pages

Best value

TOC extraction (GEPA-optimized) on Haiku 5.5

0.824≈$0.050 per 100 pages24× cheaper

  1. – sabcad-outlinerreference 0.903+0.057
  2. – sabcad-outliner+scriptreference 0.899+0.054
  3. 1 TOC extraction (sa-bcad method) 0.893+0.047
  4. 2 TOC extraction (GEPA-optimized)challenger 0.886+0.041
  5. – No skill (plain prompt) 0.846before

Tibetan · OCR models · external

Tibetan OCR

OCR engines and models ranked on BDRC's open Tibetan OCR benchmark of woodblock prints and manuscripts.

Mean character error rate, lower is better

  1. 1 BDRC OCR / Yigdzin-v1BDRC OCR 3.2%
  2. 2 MITRA-BDRC / bdrc-mitra-ocr-qwen35-0.8bMITRA-BDRC 7.3%
  3. 3 Google Vision / stableGoogle Vision 22.2%
  4. 4 Google Vision / weeklyGoogle Vision 24.7%
  5. 5 DONUT / PechaBridgeOCR_v1DONUT 28.0%

Q&A benchmark · all traditions

BuddhistBench

Questions and answers on Buddhist knowledge across traditions.

Coming soon

Scores are each task's headline metric, higher is better. Top performance is the top skill and model, whatever the price; best value is the top score among options costing at most 20% as much. Cost is the API bill for 100 pages of text (1,500 characters a page) at the models' list prices; ≈ marks estimates from runs that did not log tokens. "Mean" averages the models tested. Grey rows are reference runs: shown for comparison, not ranked, because they can't ship as the task's skill.

Coming next

How a skill gets here

  1. Workshop. Variants are built in a lab repo, such as toc-extraction-tibetan.
  2. Exam. Each one is run on the official test set and scorer in buddhist-benchmarks.
  3. Leaderboard. The result lands here, tied to an exact skill commit, model and benchmark version.
  4. Shelf. The best shippable skill is promoted to buddhist-skills.

Want to add a skill or a test set? See CONTRIBUTING.md.