[FR] Binoculars as a new metric

[MayankChaturvedi]

Binocular is a metric similar to perplexity [read about binocular]. Binocular uses two language models with same tokenization to detect the LLM generated text (or data contamination).

Currently the evaluate library doesn't include binocular in its long list of metrics. Something like the following could help the library be more useful

from evaluate import load
binocular = load("binocular", module_type="metric")
results = binocular.compute(predictions=predictions, observer_model='tiiuae/falcon-7b-instruct', performer_model='tiiuae/falcon-7b')