fugumt-ja-en-onnx

fugumt-ja-en, exported to ONNX

An ONNX export of staka/fugumt-ja-en at revision f7ce11286e1fb7a8e1f1692ff3ab68c0f9c3aecb, in the merged-decoder layout, float32. Produced for Uindosill and published so that the exact bytes its measurements describe can be downloaded and checked, rather than only reproduced.

Licence: CC BY-SA 4.0, and this export is an adaptation

The original is published under the Creative Commons Attribution-ShareAlike 4.0 International licence. Exporting a checkpoint to ONNX is adapting it, so under section 3(b) this export is Adapted Material and is licensed under the same licence, CC BY-SA 4.0, with no additional terms. The licence is linked here and from LICENSE beside these files rather than copied, which is what section 3(a)(1)(C) and 3(b)(2) ask for.

Attribution, per CC BY-SA 4.0 section 3(a)

• Creator: staka (s-taka on GitHub), the author of FuguMT. The card designates no other name, and section 3(a)(1)(A)(i) allows identification by a designated pseudonym.
• Source: https://huggingface.co/staka/fugumt-ja-en
• Copyright notice: the upstream repository was read at the pinned revision f7ce1128 on 2026-09-04 — its file listing and every text file in it — and it carries no copyright notice, so none is reproduced here rather than one being invented.
• Modifications (section 3(a)(1)(B)): the original Marian checkpoint was exported to ONNX by scripts/export-translation-onnx.py in the Uindosill repository, which splits it into an encoder graph and a decoder graph with past key values exposed behind a use_cache_branch switch. The weights are unchanged and unquantised — float32 in, float32 out. What changed is the container and the graph split. This repository redistributes the exported graphs; it does not redistribute the original checkpoint.
• Warranties: the material is provided as-is and without warranties of any kind, express or implied, to the extent permitted under section 5 of the licence.

What the source's own card says about itself, retained: "This is a translation model using Marian-NMT", source language ja, target language en, with the author's repository at https://github.com/s-taka/fugumt for details. The card reports a sacreBLEU of 39.1 on 500 randomly selected Tatoeba sentences; that is the source's figure on the source's corpus and is not repeated below as this project's.

What is here

Nine files, 378,339,764 bytes.

source.spm and target.spm are the same file: the checkpoint shares one SentencePiece model between its source and target vocabularies.

Three things that are easy to get wrong

• There is no target-language token, and one must not be added. This is a single-direction checkpoint: its vocabulary has no >>eng<<, and a source string that begins with one is tokenised as text and translated. Give it the Japanese sentence and nothing else.
• Beam search collapses at beam 6 unless the logits are re-normalised. pad_token_id, decoder_start_token_id and the single bad_words_ids entry are all token 32000. At the first decoder step the checkpoint puts almost all of its mass on that pad token; banning it leaves a distribution shifted down by a large constant, and beam search's length normalisation dilutes that constant by generating more tokens. At beam 6 猫はかわいいです。 decodes to "The Cat is slurpy slurp slurb slur slur sl sl s s". With renormalize_logits=True it decodes to "The cat is cute." Greedy decoding was never affected. Every figure below was produced with the flag on.
• generation_config.json says num_beams: 12. Every figure below was produced at beam 6, passed explicitly, batch 1, max_new_tokens=512, length penalty 1.0, early stopping off.

Measured quality, into English

FLEURS ja_jp test, all 321 distinct sentences, beam 6 as above, on one laptop CPU, 2026-09-04. chrF++ at nrefs:1|case:mixed|eff:yes|nc:6|nw:2|space:no|version:2.6.0, through the harness in the Uindosill repository (scripts/measure-translation.py).

The source-copy floor is what a hypothesis scores by echoing its own untranslated source; on Japanese it is near zero, because kanji and kana share almost no character n-gram with English, so the per-language bar Uindosill uses elsewhere is effectively a flat 45 here and passing it is a weaker statement than it is for a Latin-script source. Zero degenerate collapses and zero trailing-punctuation runs in 321 sentences. The merged layout published here reproduces the split layout's score to the digit. The export was also held to fp32 PyTorch on four Japanese sentences at beam 6 and returned the same string on four of four.

Through speech recognition: the same sentences read aloud (485 FLEURS recordings covering 287 of the 321), transcribed by parakeet-tdt_ctc-0.6b-ja at q8_0 with a character error rate of 5.54% (punctuation removed), then translated by this export, scored 48.46 — a cascade penalty of 4.01 chrF++ against a perfect transcript.

What these numbers do not cover

• No human has rated adequacy. chrF++ over read Wikipedia sentences is not adequacy, and Uindosill's own quality gate has a second, human criterion that is unperformed. That gate is not passed by its own definition.
• They are timing-free. The same 321 sentences through the same graphs on the same machine have taken between 118 s and 270 s in different runs, because the harness does not control machine load; no seconds figure from these runs describes the model.
• They are one corpus. Read speech in a quiet room, sentence by sentence; conversation, noise and long recordings are unmeasured.

Provenance

Exported with Python 3.14.7, torch 2.13.0+cpu, transformers 4.57.6, optimum 2.1.0, optimum-onnx 0.1.0, onnx 1.22.0, onnxruntime 1.29.0, sentencepiece 0.2.2. Full detail, and everything above with its caveats, is in docs/UNPROVEN.md in the Uindosill repository.