Summary

Fixes #134 by adding debug mode support to benchmarks that were missing it.

Changes

• ✅ AIME24: Added debug slicing [:2]
• ✅ AIME25: Added debug slicing [:2]
• ✅ AIW: Added debug slicing [:2]
• ✅ AMC23: Added debug slicing [:2]
• ✅ HMMT: Added debug slicing [:2]
• ✅ MATH500: Added debug slicing [:2]

Testing

# Before: These would fail or ignore debug flag
python -m eval.eval --model hf --tasks AIME24 --debug --model_args "pretrained=microsoft/DialoGPT-medium"

# After: All work with 5 examples max
python -m eval.eval --model hf --tasks AIME24,AIME25,AIW,AMC23,HMMT,MATH500 --debug --model_args "pretrained=microsoft/DialoGPT-medium"

Implementation Pattern

Following the established pattern from MTBench and other working benchmarks:
python

if self.debug:
    examples = examples[:5]

Impact

✅ Consistent debug behavior across all benchmarks
✅ Faster development iteration (5 examples vs full dataset)
✅ Reduced compute costs during testing
✅ No breaking changes to existing functionality