AKK-60m-onnx

AKK-60m-onnx

ONNX export of Thalesian/AKK-60m, 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/.

Directions

The model picks its task from an instruction placed in front of the input, joined with ": ". These are the instructions this export is registered for:

There is no instruction for transliteration back into cuneiform, so that direction is not available. The upstream card documents further "grouped" and "complex" transliteration notations, and a missing-sign infilling task, which this export does not register.

An instruction the model was not trained on does not fail. It answers anyway, in whatever direction it settles on, so the instruction is not optional.

Tokenizer

The SentencePiece model holds 32000 pieces against a vocabulary of 32518. The remainder are in added_tokens.json, and they are the cuneiform signs themselves together with the diacritics transliteration is written with. Encoding that ignores that file turns every sign into <unk> without raising.

Quality

Quality figures, training corpora and limitations are the upstream model's; read them on its card. The author reports that Akkadian to English generalises poorly outside the training domain. The context window is 512 tokens, so longer texts have to be split line by line.

The export is verified against the upstream PyTorch model rather than assumed: tokenisation matches 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 - the Akkademia project and CDLI - carry their own terms.