Describe the bug

The coqa benchmark failed silently (0 score) because it was configured with Metrics.exact_match, which is too strict for conversational QA. Additionally, the prompt function needed logic to handle one-to-many document generation (one story -> multiple Q&A pairs), which the core logic did not support efficiently.

To Reproduce

from lighteval.tasks.tasks.coqa import coqa_first_question
# Check metric config
print(coqa_first_question.metrics) # Was [Metrics.exact_match]

Expected behavior

• Metrics should be Metrics.f1_score.
• The core logic (lighteval_task.py) should support formatters returning multiple Doc objects per item.

Version info

• OS: mac
• Lighteval version: main (local development)