translate-bem-fi
Part of the Windstorm Labs open model catalogue. Scores, licence and attribution for every model: https://windytranslate.com/models/translate-bem-fi
Machine translation model, bem to fi, published by Windstorm Labs.
Attribution
Derived from Helsinki-NLP/opus-mt-bem-fi, 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 LoRA fine-tune trained on parallel corpus data and merged into the base weights.
| Method | windy-hallmark |
| Tensors modified | 36 of 254 |
| Max absolute weight delta | 7.056e-05 |
| 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.
Contents
A single transformers build in safetensors format, loadable directly:
from transformers import AutoTokenizer, AutoModelForSeq2SeqLM
tok = AutoTokenizer.from_pretrained("WindyTranslate/translate-bem-fi")
model = AutoModelForSeq2SeqLM.from_pretrained("WindyTranslate/translate-bem-fi")
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.
Evaluation
| Benchmark | chrF++ | BLEU | Band |
|---|---|---|---|
| FLORES-200 dev, 48 sentences (Bemba → Finnish) | 23.36 | 1.82 | Limited |
Measured 2026-08-04; screening score, not a publication result. Bands: Excellent ≥60, Good 45–60, Usable 30–45, Limited 15–30, Not recommended <15 (chrF++).
Limitations
- A single language direction:
bemtofi. - The evaluation above is a 48-sentence screening score, not a publication result. The weights are verified to differ from the original; that is a statement about provenance.
- Quality is not monotonic with model size or with the amount of fine-tuning applied.