EvaluationModuleError not exported from public API — sklearn errors leak to users

[YousefZahran1]

Bug

When EvaluationModule.compute() fails, raw sklearn exceptions (ValueError, KeyError) propagate to the caller with no wrapping. EvaluationModuleError exists internally but is not exported from evaluate/__init__.py, making it impossible to catch evaluate-specific errors without catching broad Exception.

Reproduce (no API key needed, evaluate 0.4.6)

import evaluate

acc = evaluate.load('accuracy')

# Test 1: raw ValueError leaks — no evaluate class in traceback
try:
    acc.compute(predictions=[], references=[])
except Exception as e:
    print(type(e).__name__)                           # 'ValueError'
    print(hasattr(evaluate, 'EvaluationModuleError')) # False

# Test 2: can't catch evaluate errors specifically
try:
    acc.compute(predictions=[], references=[])
except evaluate.EvaluationModuleError:  # AttributeError!
    pass

Expected

evaluate.EvaluationModuleError is accessible and wraps internal failures — consistent with PyTorch, sklearn, and every other major ML library.

Fix (2 lines)

evaluate/__init__.py — add:

from .module import EvaluationModuleError

evaluate/module.py ~line 467 — wrap _compute:

try:
    output = self._compute(**inputs, **compute_kwargs)
except EvaluationModuleError:
    raise
except Exception as e:
    raise EvaluationModuleError(f"Metric '{self.name}' failed: {e}") from e

Happy to open a PR for this if useful.

— Youssef Ibrahim | github.com/YousefZahran1
(Previously: #753 — added zero_division to F1 metric)