Document unconditioned_query requirement for PMI normalization

src/lighteval/metrics/normalizations.py

@@ -471,6 +471,21 @@ def _inner_normalizer(text: str) -> str:
 class LogProbPMINorm:
     """Performs Pointwise mutual information normalization. log_likelihood_conditioned - log_likelihood_unconditioned.
     Useful when answer contains generally unlikely tokens.
+
+    This normalization requires the task's `prompt_function` to set `Doc.unconditioned_query` to a
+    version of the prompt that contains no task-specific context (e.g. an empty string, or just
+    `"Answer:"`), so that the unconditioned log-likelihood of each choice can be computed:
+
+    ```python
+    Doc(
+        query=f"Question: {question}\\nAnswer:",
+        choices=choices,
+        gold_index=gold_index,
+        unconditioned_query="Answer:",
+    )
+    ```
+
+    If `Doc.unconditioned_query` is not set, this normalization cannot be applied.
     """
 
     name: str = "norm_pmi"