Sphragis benchmark leaderboard

Benchmark results

Sphragis

1 model

Select a row with diagnostics to compare accuracy by author

			Sentence tasks				Source-held-out tasks

OLMo-1B authorial language models	1.18B parameters per author	Lowest per-token perplexity	62.36	86.84	89.53	92.44	57.80	69.88	67.59	56.72	Per-author base model and epoch selected on validation attribution; the source-held-out track has its own models, trained on its own split

Scores are held-out test macro-F1 (%).

Benchmark results

1 model

Select a row with diagnostics to compare accuracy by author

			Verse tasks

OLMo-1B authorial language models	1.18B parameters per author	Lowest per-token perplexity	56.81	76.15	80.99	72.88	Seventeen models trained on verse_1; per-author base model and epoch selected on validation attribution

Scores are held-out test macro-F1 (%).