Referenceless QE metrics require reference input

[zouharvi]

🐛 Bug

When using comet from the evaluate package, it seems to require the reference also for QE metrics. There's an easy but dirty hotfix.

To Reproduce

comet_metric = evaluate.load("comet", config_name='wmt21-comet-qe-da')
# erorr!
comet_metric.compute(predictions=["Hallöchen"], sources=["Heyo"])

More info

Passing a reference with an empty string appears to satisfy the library. I checked that the output scores are not dependent on the reference:

# {'mean_score': 0.08140071481466293, 'scores': [0.08140071481466293]}
comet_metric.compute(predictions=["Hallöchen"], sources=["Heyo"]. references=[""])
# {'mean_score': 0.08140071481466293, 'scores': [0.08140071481466293]}
comet_metric.compute(predictions=["Hallöchen"], sources=["Heyo"], references=["Howdy"])

[ricardorei]

Hi @zouharvi, This seems to be an issue on the evaluate package and not on unbabel-comet. I suggest you open and issue there...