cuneiformBase-400m-onnx

ONNX export of Thalesian/cuneiformBase-400m, for CPU inference with linguonnx. The weights are unchanged; this repository adds the exported graphs and an int8 quantisation.

Both precisions ship the three-graph encoder/decoder/decoder-with-past split that linguonnx drives: fp32 at the top level, int8 under int8/.

Languages

Akkadian, Sumerian, Hittite and Linear B (Mycenaean Greek), to and from English, and Hittite to and from German. Each ancient language is carried in two written forms — the signs themselves and a Latin transliteration — which this export treats as separate languages, because a model that reads one does not necessarily read the other.

Elamite is in the upstream training data but has no published evaluation and no documented instruction, so it is not claimed here.

Directions

The model picks its task from an instruction placed in front of the input, joined with ": ". An instruction the model was not trained on does not fail: it answers anyway, in whatever direction it settles on. The full set this export registers is in linguonnx's registry entry; the upstream card lists them all, including "simple" and "complex" transliteration notations that this export does not register separately.

Tokenizer

This export has no SentencePiece model. Its whole vocabulary — a Unigram model of 256300 pieces plus 1089 added tokens — is in tokenizer.json. linguonnx reads that file directly with its own Unigram implementation rather than depending on the tokenizers package.

Note that AutoTokenizer cannot load the upstream repository: its tokenizer_config.json gives extra_special_tokens as a list where transformers expects a mapping. Reading tokenizer.json with the tokenizers library directly works, and is what the export was verified against.

Quality

Quality figures, training corpora and limitations are the upstream model's; read them on its card. The author reports that out-of-domain Akkadian generalises poorly, and scores lower on Akkadian than their earlier Akkadian-only models. The context window is 512 tokens, so longer texts have to be split line by line.

The export is verified against the upstream model rather than assumed: tokenisation matches the reference token for token, and beam-search output matches string for string across every direction above.

Licence

Apache-2.0, following the upstream model. The corpora behind it — CDLI, the Akkademia project, HPM and the published Linear B resources — carry their own terms.