Hi!

For the current generation of NLLB models (and the next generation is not expected in the foreseeable future), there is no obvious way to softly prompt the model towards specific domains or specific translations.

The obvious things that you could try are:

1. Fine-tune NLLB with an English-Chinese dataset in the sports domain (you'll need to find the data first, of course).

2. For a specific translation, if you want to force it to contain a specific word, you can use positively constrained beam search algorithms, like the one presented in https://huggingface.co/blog/constrained-beam-search.