Ive wanted to use Argos for some help in Translating a few things for my work and installed libre for testing purposes offline on a Linux VM. And holy shit whats going on with the accuracy ? here is a real life Example:
werkstueckmessen bei offener Werkstueckspannung nicht moeglich, Spannbacken am Reitstock offen
---->workbench measurements at open workbench voltage not possible, clamping baking on the riding stock open
Do i have to train it myself or will this get better ?
Are there any Workarounds beside using DeepL ?
With the model i have trained, i am getting the following translation:
The German French package doesn’t yield as great a result so i guess the relevance of translations is due to the Europat dataset which is available in German English but not German French. But then, i have worked the dataset out in such a way that performance is way better than what you would get if training with the raw corpora. Plus I have researched transformer hyperparameters extensively to translate in a syntactically accurate way between German and English and French, tweaked the ctranslate2 dependancy, Locomotive’s train.py too…
Those models are not published yet, but they should be open-sourced in the next few months.