Description
In recent years, "evaluation" has taken a specific meaning in the language of foundation models, more aligned with LMEval, LLM-as-a-Judge, etc. It no longer means to perform inference from a model.
There as also significant confusion because Evaluator.evaluate
clashes with the Metric.evaluate(...)
function. The latter usage is correct.
Given this, it makes sense to rename Evaluator
subclasses to Predictor
and the .evaluate(...)
method to .predict(...)
.
Potential other names considered:
InferenceManager
: too long (annoying while using in notebook). "Manager" is a meaningless term here which does not convey the semantics.Inference
: makes sense as a subclass (LocalInference
,RayInference
, etc) but not asInferrer
: technically correct but not a real word.