Add FEVER Metric for Fact Verification Evaluation

[mkhi238]

Description

This PR adds the FEVER (Fact Extraction and VERification) metric to the evaluate library.

What does this metric do?

The FEVER metric evaluates fact verification systems against evidence retrieved from Wikipedia. It includes:

• Label accuracy: Measures how often predicted labels match gold labels
• FEVER score: The official metric - requires both correct label AND complete evidence retrieval
• Evidence precision/recall/F1: Micro-averaged metrics for evidence retrieval

Checklist

• Implementation in fever.py
• Comprehensive test suite with 12 tests (all passing)
• README.md with metric card and examples
• Gradio app for interactive testing
• Code formatted with black and isort
• All tests pass

Tests:

All tests pass successfully:

test_fever.py
............
----------------------------------------------------------------------
Ran 12 tests in 0.142s

OK

References

• FEVER Paper: https://arxiv.org/abs/1803.05355
• FEVER Dataset: https://fever.ai/dataset/

Note: This is my first contribution to the evaluate library. I'd appreciate any feedback or suggestions for improvement!

[lhoestq]

lgtm !

[lhoestq]

the metric is available now, with its page at https://huggingface.co/spaces/evaluate-metric/fever :)