Urdatorn's picture
Publish rebuilt atomic model benchmarks
4b99898
|
Raw
History Blame
1.5 kB
metadata
title: Sphragis Leaderboard
emoji: 🥇
colorFrom: green
colorTo: indigo
sdk: static
app_file: index.html
pinned: true
license: apache-2.0
short_description: Leaderboard for Ancient Greek author attribution
fullWidth: true
datasets:
  - Urdatorn/sphragis
tags:
  - leaderboard
  - authorship-attribution
  - ancient-greek

Sphragis Leaderboard

A statically generated SvelteKit + TypeScript leaderboard for the current Sphragis Ancient Greek authorship-attribution benchmarks. The UI uses Svelte 5 runes and runs without Python, CPU, or GPU Space hardware.

The rebuilt leaderboard's typed task catalog in src/lib/data.ts covers:

  • Sphragis sentence_{1,5,10,50}
  • Sphragis source_held_out_{1,5,10,50}
  • Sphragis-Metre verse_{1,5,10,50}

Published contenders include OLMo-1B authorial language models, character n-grams, end-to-end and frozen GreBERTa, lemma TF-IDF, Burrows lemma z-scores, syntax feature rates, and (for Sphragis-Metre) metrical line features. All applicable feature orders and dimensions are selected independently for every task using validation macro-F1. Scores, per-author accuracies, and confusion matrices come from the same saved held-out predictions as the reported score.

Install and run locally:

npm ci
npm run dev

The table interaction and visual architecture adapt patterns from the Apache-2.0-licensed embeddings-benchmark/leaderboard-frontend. See NOTICE for attribution.