[MonikaDvorackova]

Summary

Addresses #674.

• Replace eager package-level imports with lazy loading.
• Avoid importing heavy dependencies such as datasets, pandas, and pyarrow during a plain import evaluate.
• Preserve package-level API exports, including the callable evaluate.evaluator.
• Add import benchmarks and regression tests.

Benchmark

Representative measurements on macOS:

Import	Before	After
import evaluate	14.914 s	0.119 s
from evaluate import load	14.099 s	7.238 s
from evaluate import EvaluationSuite	9.995 s	15.325 s
from evaluate import evaluator	11.657 s	11.220 s

Import timings varied substantially between runs. The main improvement is consistently observed for plain import evaluate; other imports may still load heavy dependencies on first access.

Validation

• 11 regression tests passed.
• Ruff lint and formatting checks passed.
• Verified 50 public exports and evaluator compatibility.

The complete test suite was not run because some optional dependencies were not installed.

Additional local benchmark (Python 3.10, macOS)

Seven subprocess measurements per import; median values.

Import	Upstream main	PR branch
import evaluate	1.419 s	0.024 s
from evaluate import load	1.563 s	0.768 s
from evaluate import EvaluationSuite	1.456 s	1.389 s
from evaluate import evaluator	1.313 s	1.528 s

Plain import evaluate was approximately 59x faster.
The evaluator import was slower in this measurement.
Results may vary across runs and environments.

Additional validation:

• 6 focused lazy-import tests passed.
• All 50 original public names remained accessible.
• Wildcard exports were verified.
• Regression tests added in commit b6c1136.

The complete project test suite has not yet been verified in this environment.