fix(normalizations): guard against index out of range in LogProbToken…

[inakiLakunza]

…Norm

Problem

When running evaluations with cached predictions, normalize_log_probs crashes with an IndexError: list index out of range inside the LogProbTokenNorm case. This happens because the cached output_tokens list can be shorter than choices_logprob, for example when a task's choices change between runs but the cache is not invalidated.

Affected path:
lighteval/metrics/normalizations.py → normalize_log_probs → LogProbTokenNorm branch

Fix

• Use min(len(choices_logprob), len(choices_tokens)) to safely cap the iteration range instead of blindly iterating over len(choices_logprob).
• Emit a logger.warning when truncation occurs so users are alerted to potential cache corruption and can take action (e.g. clearing the cache).
• Add a module-level logger = logging.getLogger(__name__) for consistent logging.

Changes

• src/lighteval/metrics/normalizations.py
  • Added import logging and module-level logger
  • Replaced bare list comprehension in LogProbTokenNorm with length-guarded version

How to reproduce the original bug

1. Run an evaluation that uses LogProbTokenNorm (e.g. belebele_mkd_Cyrl_cf).
2. Allow results to be cached.
3. Modify the number of choices for the task, or use a stale cache from a previous run.
4. Re-run, the pipeline crashes at the metric computation stage with IndexError: list index out of range.