A Marian translator nudged toward literary prose with a LoRA adapter, merged and quantized to run on your own machine. Nothing is sent to a server.
Downloading the model (about 140 MB, cached after the first visit)…
The base model is opus-mt-en-es, already a capable translator. A LoRA adapter was trained on opus_books, a corpus of translated literature, to shift it toward that register. The adapter was then merged into the weights and exported so it can run here.
| Base model | 17.09 sacreBLEU |
|---|---|
| With the adapter | 19.99 sacreBLEU |
| Evaluated on | 1,000 held-out pairs |
That gain is real but diffuse. On most sentences the two models produce identical output; the difference shows up as small changes in word choice and phrasing. If you are comparing against another translator, expect the gap to be subtle.
Tuned on translated literature from the 19th and early 20th centuries. Technical documentation, conversational messages and contemporary slang are outside what it saw, and quality there is unmeasured.
The int8 quantization that makes this small enough to download introduces small differences from the full-precision model. One direction only: English to Spanish.
It translates; it does not check whether the result is appropriate for your context. For anything consequential, have a Spanish speaker read it.