Summary

EvaluationModuleError did not exist in the public API, so users had no way to catch evaluate-specific errors — raw ValueError / KeyError from sklearn or numpy leaked straight to the caller.

Changes

src/evaluate/module.py

• Introduces EvaluationModuleError(Exception) right after the logger setup
• Wraps the self._compute(**inputs, **compute_kwargs) call in compute() so any unexpected exception is re-raised as EvaluationModuleError with a descriptive message; already-typed errors are re-raised as-is

src/evaluate/__init__.py

• Exports EvaluationModuleError alongside EvaluationModule

Before / After

import evaluate

acc = evaluate.load("accuracy")

# Before:
# hasattr(evaluate, "EvaluationModuleError")  # False
# acc.compute(predictions=[], references=[])  # raises raw ValueError

# After:
print(hasattr(evaluate, "EvaluationModuleError"))  # True

try:
    acc.compute(predictions=[], references=[])
except evaluate.EvaluationModuleError as e:
    print(f"Caught: {e}")
# Caught: Metric computation failed for 'accuracy': ...

Fixes #758