I have been using a Custom translator for 3 years — the task I'm doing: translating program code from Russian to English. Every year I train a new model for translation, as my data for training the model is supplemented and extended.
Last week I trained another translator model and faced the following problems:
- Training duration. The training took 1 day and 4 hours. Last year the model was trained in 4 hours. (The amount of data to train increased by about 10%.)
- Quality of translation. The BLEU score came out to be 62.06 (baseline BLEU 16.98). Last year, the translation model got a BLEU of 96.19 (baseline BLEU 23.34) on the same test dataset.
In fact, many of the translations produced by the model contradict the data used in training.
My questions:
- What could the degradation in translation quality be related to?
- Is it now possible to choose the previous year's version of the Custom translator for training?