How this translator works
You type Old Assyrian in standard transliteration (hyphens, Sumerograms such as KÙ.BABBAR, Kaneš trade jargon). This translator targets the merchant-archive dialect behind the Living Rosetta simulation, not later Babylonian.
ByT5 (Google’s byte-level T5) runs as an ensemble of separately fine-tuned checkpoints on the same line. Each copy proposes several beam hypotheses; scores are length-normalized, and one line wins across the whole ensemble. That draft is what you see in the Lab Console before any polish.
A lexicon pass and retrieval still anchor rare signs to attested forms. When the API offers it, an optional LLM polish pass rewrites the winning English for readability while keeping the tablet sense.
- Bytes in, bytes out — no word-piece tokenizer between you and cuneiform transliteration.
- Ensemble + beams reduce single-model blind spots on damaged or formulaic lines.
- The green status dot reflects whether the remote model stack answered; gloss-only mode still works offline.