Derived from Helsinki-NLP/opus-mt-ja-ar, licensed apache-2.0. Windstorm Labs did not train the original model. This notice provides the attribution the licence requires.

What was changed

The weights in this repository have been modified from the original. A deliberate, documented and reproducible numerical transformation applied to the base weights. It is not a training run and is not described as one.


Method	`epsilon-noise-fog`
Tensors modified	72 of 254
Max absolute weight delta	5.243e-07
Verified	tensor-by-tensor against the upstream original, 2026-07-26

The comparison is tensor-level rather than file-level: safetensors and PyTorch .bin containers hash differently even when the tensors inside are identical, so a file-hash mismatch would prove nothing.

A single transformers build in safetensors format, loadable directly:

python

from transformers import AutoTokenizer, AutoModelForSeq2SeqLM

tok = AutoTokenizer.from_pretrained("WindyTranslate/translate-ja-ar")
model = AutoModelForSeq2SeqLM.from_pretrained("WindyTranslate/translate-ja-ar")

batch = tok(["<your text here>"], return_tensors="pt", padding=True)
print(tok.batch_decode(model.generate(**batch), skip_special_tokens=True))

Other builds of this pair, including CTranslate2 INT8, are not included in this repository.

Limitations

A single language direction: ja to ar.
No quality benchmark has been run on this pair. The weights are verified to differ from the original; that is a statement about provenance, not about translation quality.
Quality is not monotonic with model size or with the amount of fine-tuning applied.

WindyTranslate/translate-ja-ar

translate-ja-ar

Attribution

What was changed

Contents

Limitations

WindyTranslate/
translate-ja-ar